bunch of tokenization related bug-fixes

[molereddy]

What does this PR do?

Resolves several bugs related to narrow tokenization cases. 2, 3, 4 are interlinked issues and 1 is close to them as well.

1. Log-probs

• Previously the tokenwise_logprobs and tokenwise_vocab_logprobs functions were inadvertently keeping the eos token in the label sequence, leading to the model being forced to predict eos token in the MIA (minK, minK++, gradnorm) and ES/EM metrics.
• This wasn't much of an issue w TOFU as the chat model predicts eos anyways, but still the original intention was to never get eos involved in these calculations.
• This became especially an issue with MUSE (which is not a chat model thus doesn't predict eos) leading to low ES scores issues (Extraction Strength on knowmem(MUSE) and perturb(TOFU) #100)
• This PR fixes the code, the results and existing evals are to be updated accordingly

2. Small target sequences getting mixed up w prompt

• For some models like Phi-1.5 which have weird tokenizers and for data points which have 1 word target labels, you may end up with the target words being tokenized into the prompt
• This PR warns about such cases and modifies ES and EM to set value as None in such cases

3. Handing empty label sequence cases

• When the above issue occurred, previously the code did not update both the labels and logprobs list, leading to mismatch and mixup across the batch which then led to failed asserts with Phi-1.5 [pointed out in Extraction Strength on knowmem(MUSE) and perturb(TOFU) #100]
• This PR ensures empty tensors are sent in such cases

4. Date in system prompt

• Llama3.2 or 3.1 tokenizer inexplicably adds the current date to the chat template prompt, making results (slightly) irreproducible (Reproduction issues on LLaMA3.1 (TOFU): checkpoint vs final‑model evaluation + epoch step‑count mismatch #88)
• This PR allows for fixing the date for Llama through the model template args

5. Handling invalid values

• When a metric is evaluated to be None on a data point (usually due to tokenization issues), we set the aggregation to filter out such points and only compute on those that have values

6. Misc modifications

• The tokenwise_logprobs and tokenwise_vocab_logprobs involved changing model train mode. This was better handled in the evaluation metric code itself (used only by gradnorm MIA attack).
• Documentation fix to remove non-existent cfg file

Fixes # (issue)
#98
#100

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Have you gone through the contributions guide?
• Are your changes documented? Read documentation guidelines here.