[Bugfix] Only quant-compress modules with weight quantization

[kylesayrs]

Purpose

• Skip compression for attention modules, which have a quantization config but do not have weights to quantize
  • Without these changes, attention modules will be compressed. This is not a problem in itself (since they have no weights, they are no opped), but [Bugfix] Safeguard against submodule parameter deletion in decompress_model #347 fixes a bug which occurs when non-leaf modules are attempted to be compressed

Changes

• map_module_to_scheme used to only quantize compress leaf modules. Now, this method only quantize compresses modules with weight quantization

Testing

• KV cache tests pass with [Bugfix] infer_quantization_format when model only has activation quantization vllm-project/llm-compressor#1635