[CUDA/CPU/WebGPU] Add state_window to LinearAttention and CausalConvWithState

[tianleiwu]

Description

Adds an optional state_window attribute (default 0 = current behavior) to the
LinearAttention and CausalConvWithState contrib operators, implemented for CPU, CUDA and
WebGPU.

When state_window = W > 0, the recurrent state tensors gain a leading window axis:

op	state shape, state_window = 0	state shape, state_window = W
CausalConvWithState	(batch_size, channels, kernel_size - 1)	(W, batch_size, channels, kernel_size - 1)
LinearAttention	(batch_size, num_kv_heads, head_size_k, head_size_v)	(W, batch_size, num_kv_heads, head_size_k, head_size_v)

Slot W-1 holds the state after the last token of the step; slot j holds the carry state after
the j-th token. W = 0 keeps the existing 3D/4D shapes, so this is backward compatible and no
existing model changes behavior.

Motivation and Context

These two ops carry a recurrent state, which makes them the blocker for multi-token
(speculative / MTP) decoding. A full-attention model can verify N+1 draft tokens in one forward
and then simply discard the rejected KV-cache tail. A linear-attention model cannot: after a
verify step the state reflects all N+1 tokens, so if only the first k are accepted there
is no way to recover the state at position k — the only correct fallback is to re-run the
recurrence over the accepted prefix.

With state_window, the op writes the intermediate per-token carry states into the window as it
sweeps, so the rejection path is a slot select instead of a recomputation. This is what allows
num_speculative_tokens = 3 (verify batch M = 4, recurrent_state_window = 4) to run on
Qwen3.6-35B-A3B, a hybrid model with 30 linear-attention layers out of 40.

This PR is enablement, so it has no standalone kernel speedup to report. For context, the
end-to-end payoff it unlocks, measured on 1x H200 SXM (SM90, CUDA 13.0) with
Qwen3.6-35B-A3B-NVFP4 + MTP N=3, CUDA graph on:

context	ms/step	decode tok/s	tokens accepted / step
256	8.441	325.0	2.744
16384	9.567	283.9	2.716

Also included: a comment documenting that GQA's FlashDecoding fast-decode path is valid for
sequence_length >= 1 (its causal masking and split-KV reduction match regular FlashAttention),
plus multi-token decode coverage in test_gqa.py.

Tests

• onnxruntime/test/contrib_ops/causal_conv_with_state_op_test.cc — new windowed cases.
• onnxruntime/test/contrib_ops/linear_attention_op_test.cc — new windowed cases.
• onnxruntime/test/python/transformers/test_gqa.py — multi-token (MTP-shaped) decode.

Windowed and non-windowed runs are compared against the same reference, so state_window = W
must reproduce slot-by-slot what W = 0 produces when stepped one token at a time.

Note

docs/ContribOperators.md still needs to be regenerated for the two schema changes
(python tools/python/gen_contrib_doc.py after a build). Keeping this as a draft until that
is in. It also touches group_query_attention.cc / bert_defs.cc / test_gqa.py, which
overlap #29904 — whichever lands second will need a rebase.