Make AutoencoderKLWan stateless to fix torch.compile compatibility

[NotNANtoN]

What does this PR do?

Fixes #14770

This PR makes AutoencoderKLWan (used in Wan2.1 and Wan2.2 models) stateless during execution so that it compiles cleanly with torch.compile without graph breaks or issues from mutating internal module attributes.

Key Changes

1. Stateless Feature Caching: Allocate feat_map = [None] * self._cached_conv_counts[...] locally inside _encode, _decode, tiled_encode, and tiled_decode and pass it down along with a local conv_idx = [0] list.
2. Eliminate Mutable Defaults: Changed mutable default arguments from feat_idx=[0] to feat_idx=None in all residual / up / down / mid blocks.
3. Backward Compatibility: clear_cache() is preserved as a method on AutoencoderKLWan so existing external scripts calling vae.clear_cache() do not break.

Verification & Performance

Tested on NVIDIA RTX 4090 (PyTorch 2.11 / CUDA 12.8, BF16, 5 frames 480x640):

• Numerical Parity: max_abs_diff = 0.0 (exact bitwise parity with eager baseline).
• Unit Tests: All 32 unit tests in tests/models/autoencoders/test_models_autoencoder_wan.py pass cleanly.
• Latency Benchmarks:
  • Eager uncompiled: 84.43 ms
  • torch.compile(mode="default"): 61.10 ms (1.38x speedup)
  • torch.compile(mode="max-autotune-no-cudagraphs"): 58.38 ms (1.45x speedup)

Self-Review Notes

• Diff Scope: Confined strictly to src/diffusers/models/autoencoders/autoencoder_kl_wan.py.
• Formatting: Verified with ruff check and ruff format.
• Edge cases: Verified both standard encoding/decoding and tiled encoding/decoding (test_tiling_encode_shape, test_tiling_decode_shape).

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR?
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Did you read our philosophy doc?
• Was this discussed/approved via a GitHub issue or the forum? Fixes [Bug / Optimization] Make AutoencoderKLWan stateless to fix torch.compile compatibility and enable 1.45x speedup #14770
• Did you make sure to update the documentation with your changes? (Internal implementation fix, no public API changes)
• Did you write any new necessary tests? (Existing comprehensive test suite covers all modified blocks)
• Are you the author (or part of the team) of the model/pipeline?

Who can review?

@yiyixuxu @dg845 @asomoza