feat: Better support for dynamic forcings and coupling (ecmwf#262)

## Description

Some refactoring of inference to allow the clean split between inputs:
initial conditions, forcings and constants, so each category can be
loaded from a different source.

### Variables categories


For that we first categorise the variables into:

 - computed
 - forcing
 - diagnostic
 - prognostic
 - constant
 - accumulation

A variable can belong to several categories, i.e. `computed` and
`constant`.

The runner now selects variables by categories using the new
`Checkpoint.select_variables()`, which leads to a cleaner and much
shorted code.


### Separate inputs

The runner has now three inputs:

- `input`: used to fetch the prognostics for the initial conditions
(e.g. `2t` in an atmospheric model).

- `constant_forcings`: used to fetch the constants for the initial
conditions (e.g. `lsm` or orography)

- `dynamic_forcings`: used to fetch the forcings needed be some models
throughout the length of the forecast (e.g. atmospheric fields used as
forcing to an ocean model)


To ensure backward compatibility, unless given explicitly in the config,
`constant_forcings` and `dynamic_forcings` both fallback to the `input`
entry.

A new config option lets the user select which category of variables are
written to the output if `write_initial_condistions` is `true`. For
backward compatibility, it defaults to `prognotic` and
`constant_forcings`


### New commands

`anemoi-inference inspect --variables checkpoint.ckpt`

Will list the training variables and their categories.

`anemoi-inference inspect --datasets checkpoint.ckpt`

Will print the arguments passed to anemoi-dataset during training.

`anemoi-inference inspect --requirements checkpoint.ckpt`

Will print a Python's `requirements.txt` based on the versions of the
packages used during training.

## What problem does this change solve?

<!-- Describe if it's a bugfix, new feature, doc update, or breaking
change -->

## What issue or task does this change relate to?

Closes ecmwf#245 

##  Additional notes ##
<!-- Include any additional information, caveats, or considerations that
the reviewer should be aware of. -->

***As a contributor to the Anemoi framework, please ensure that your
changes include unit tests, updates to any affected dependencies and
documentation, and have been tested in a parallel setting (i.e., with
multiple GPUs). As a reviewer, you are also responsible for verifying
these aspects and requesting changes if they are not adequately
addressed. For guidelines about those please refer to
https://anemoi.readthedocs.io/en/latest/***

By opening this pull request, I affirm that all authors agree to the
[Contributor License
Agreement.](https://github.com/ecmwf/codex/blob/main/Legal/contributor_license_agreement.md)

BEGIN_COMMIT_OVERRIDE
feat: Better support for dynamic forcings and coupling in prepml (ecmwf#262)
feat: Add requirements subcommand (ecmwf#245)
feat: Lazy torch loading (ecmwf#277)
END_COMMIT_OVERRIDE


<!-- readthedocs-preview anemoi-inference start -->
----
📚 Documentation preview 📚:
https://anemoi-inference--262.org.readthedocs.build/en/262/

<!-- readthedocs-preview anemoi-inference end -->

---------

Co-authored-by: Gert Mertes <gert.mertes@ecmwf.int>
Co-authored-by: Harrison Cook <Harrison.cook@ecmwf.int>
Co-authored-by: Matthew Chantry <chantry.m.j@gmail.com>
Co-authored-by: bdvllrs <benjamin.devillers@meteo.fr>

Author: 5 people

.gitignore

@@ -141,3 +141,5 @@ _version.py
 *.to_upload
 tempCodeRunnerFile.python
 Untitled-*.py
+data
+tmp/

docs/cli/inspect.rst

@@ -1,4 +1,4 @@
-.. _inspect_command:
+.. _inspect-command:
 
 Inspect Command
 ===============

docs/cli/introduction.rst

@@ -13,9 +13,9 @@ The tools can provide help with the ``--help`` options:
 
 The commands are:
 
-- :ref:`Run Command <run_command>`
-- :ref:`Metadata Command <metadata_command>`
-- :ref:`Inspect Command <inspect_command>`
-- :ref:`Validate Command <validate_command>`
-- :ref:`Patch Command <patch_command>`
-- :ref:`Requests Command <requests_command>`
+- :ref:`Run Command <run-command>`
+- :ref:`Metadata Command <metadata-command>`
+- :ref:`Inspect Command <inspect-command>`
+- :ref:`Validate Command <validate-command>`
+- :ref:`Patch Command <patch-command>`
+- :ref:`Requests Command <requests-command>`

docs/cli/metadata.rst

@@ -1,4 +1,4 @@
-.. _metadata_command:
+.. _metadata-command:
 
 Metadata Command
 ================

docs/cli/patch.rst

@@ -1,4 +1,4 @@
-.. _patch_command:
+.. _patch-command:
 
 Patch Command
 =============

docs/cli/requests.rst

@@ -1,4 +1,4 @@
-.. _requests_command:
+.. _requests-command:
 
 Requests Command
 ================

docs/cli/run.rst

@@ -1,4 +1,4 @@
-.. _run_command:
+.. _run-command:
 
 Run Command
 ===========

docs/cli/validate.rst

@@ -1,4 +1,4 @@
-.. _validate_command:
+.. _validate-command:
 
 Validate Command
 ================

docs/index.rst

@@ -128,6 +128,7 @@ You may also have to install pandoc on MacOS:
    :hidden:
    :caption: User Guide
 
+   inference/input-types
    inference/parallel
    inference/apis/introduction
    inference/configs/introduction

docs/inference/apis/level3.rst

@@ -5,7 +5,7 @@
 ##################
 
 You can run the inference from the command line using the
-:ref:`anemoi-inference run <run_command>` command.
+:ref:`anemoi-inference run <run-command>` command.
 
 You must first create a configuration file in YAML format. The simplest
 configuration must contain the path to the checkpoint:
@@ -66,7 +66,7 @@ checkpoint that was trained with one the ICON grid:
 .. literalinclude:: code/level3_4.yaml
    :language: yaml
 
-See :ref:`run_command` for more details on the configuration file.
+See :ref:`run-command` for more details on the configuration file.
 
 .. warning::