Add --config_source pretrained to capture full model config via model_id

[kiszk]

Summary

Adds a --config_source option to auto_generate_module_config.py that
controls how a module's PretrainedConfig constructor argument is recorded in
the generated YAML.

• kwargs (default, unchanged): rebuild the config from a handful of extracted
  dimensions (config_path + config_kwargs). Library defaults fill in
  everything else — the historical behaviour.
• pretrained (new): record only the model_id so the OOT test framework can
  load the full, faithful config with AutoConfig.from_pretrained(model_id).

Motivation

The kwargs path extracts only ~6 fields (hidden_size,
num_attention_heads, num_key_value_heads, intermediate_size,
max_position_embeddings, _attn_implementation). Every other field
(rope_theta, rope_scaling, attention_bias, head_dim, logits_scaling,
...) is dropped to library defaults, so a module rebuilt from the YAML can
diverge from the real model. Recording the model_id and reloading the config
at test time preserves every field the model actually had.

There are two other motivations for this PR (out of scope of this PR)

• Will support calling LLM() for instead of instantiating config_cls to support this architecture
• Will introduce opportunities to apply something (e.g. monkey patches and optimizations of device layouts of weight tensors beyond generic layout) at a model loading time.

Changes (utils/module_discovery/auto_generate_module_config.py)

• New CONFIG_SOURCE_KWARGS / CONFIG_SOURCE_PRETRAINED constants and a shared
  _build_config_arg_spec(config, config_source, model_id) helper that produces
  the config-arg spec, replacing the two inlined spec-building blocks in
  capture_constructor_info.
• In pretrained mode the spec emits model_id plus a small config_overrides
  carrying the resolved _attn_implementation (which from_pretrained may
  leave as None), keeping the standalone module on the same attention path
  used at capture time. config_path is still recorded as a human-readable
  hint.
• ModuleInfoCapture.__init__ gains config_source / model_id (validated),
  so switching works both from the CLI and from the driver API used by
  generate_*_config.py scripts.
• _convert_constructor_arg_to_sample_input emits the model_id /
  config_overrides form when present, else the existing config_kwargs form.
• New --config_source {kwargs,pretrained} CLI flag (default kwargs); main
  passes it plus model_id=args.model_path into ModuleInfoCapture.

Fully backward compatible: the default (kwargs) reproduces the previous output
byte-for-byte, and existing generated YAMLs are unaffected.

Related

This is the generator half of the change. The paired consumer change (the OOT
framework resolving model_id via AutoConfig.from_pretrained) is submitted
separately against the torch-spyre repo.

#173

Testing

• Verified the spec logic in isolation: kwargs and pretrained produce the
  expected shapes, pretrained without a model_id raises, and
  ModuleInfoCapture rejects an invalid config_source.
• Usage:
  python utils/module_discovery/auto_generate_module_config.py --model_path ibm-granite/granite-3.3-8b-instruct --config_source pretrained -o /tmp/granite_pretrained.yaml
• Verified end-to-end run through the OOT runner on the Spyre