feat: capture decode-step KV cache in auto-generated module configs

[kiszk]

What this PR does

The auto-generated module-test configs drive each DecoderLayer through both a
prefill and a decode invocation, but the forward hook was skipping
past_key_values entirely:

if key in ("past_key_values", "past_key_value"):
    continue  # Skip - not needed for module-level tests

As a result the decode invocation was replayed with past_key_values=None. In
HF attention the cache-update branch

if past_key_values is not None:
    key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

is then never taken, so the "decode" step silently degrades to a 1-token
self-attention (Q[1] × K[1]) instead of attending over the cached past.
Both eager and compiled runs share the same wrong input, so test_eager_vs_compile
passes without actually exercising decode.

This PR captures the decode-step KV cache so the module test can rebuild an
equivalent cache and drive the real "attend over past + new token" path. The
companion torch-spyre change reconstructs the cache on the test side.

Changes (utils/module_discovery/auto_generate_module_config.py)

• _resolve_layer_idx(module) — resolves the decoder layer index. In
  transformers ≥5 layer_idx lives on self_attn, not the DecoderLayer, so
  we fall back through both.
• _extract_cache_info(...) — snapshots the target layer's populated K/V slice
  from a StaticCache into a cache spec (cache_path, layer_idx,
  key/value, max_cache_len, config_path, config_kwargs):
  • Uses the per-layer get_seq_length(), not the whole-cache length
    (which over-counts by one at layer i>0 once earlier layers have updated
    this pass).
  • Records only the populated [:, :, :past_len, :] slice (StaticCache
    allocates the full max_cache_len up front).
  • Only StaticCache is recorded; a growable DynamicCache has no fixed
    shape to reconstruct, so it emits a warning and is skipped.
  • Prefill (empty slot) records nothing — equivalent to past_key_values=None.
• Hook no longer skips past_key_values; it records the cache spec instead.
• _convert_captured_input_to_sample_input / _extract_pattern gain a cache
  branch (the latter keeps prefill and decode as distinct invocation signatures).
• _validate_cache_mask_consistency(...) — generation-time sanity check that a
  cached decode invocation also carries an attention_mask whose key length can
  cover the cached past (logged, non-fatal).

Verification (CPU, transformers 5.12)

• Ran the capture path on a tiny GptOssForCausalLM (same modeling code as
  gpt-oss-20b, incl. past_key_values.update()) driven with an explicit
  StaticCache; the written YAML's GptOssDecoderLayer decode invocation
  contains the cache spec (StaticCache, layer_idx=0, key
  [1, 2, 16, 16] = [B, num_kv_heads, past_len, head_dim]).
• End-to-end: capture → reconstruct StaticCache (seq_length 16) → real
  DecoderLayer decode forward runs update() over past 16 + new 1 token
  (output [1, 1, 64], cache length 16 → 17).
• DynamicCache still warns and records no cache.

Notes for reviewer

• Requires the companion torch-spyre PR
  which adds the cache spec reconstruction to the OOT framework.
• The new generate_*_config.py drivers assume the reusable
  capture_module_invocations / load_model_only / write_module_config API which Refactor auto_generate_module_config: extract reusable forward-pass API from main() #131 will
  introduce. So the driver is not runnable here until Refactor auto_generate_module_config: extract reusable forward-pass API from main() #131 lands.