Add KV cache to the EAGLE-3 draft head

[digantdesai]

Adds a flat KV cache and an explicit-mask attention path (forward_cached) to the
draft head so a proposal step reuses cached keys/values instead of recomputing
the prefix's projections and MLP. Attention still scores against the full
max_seq_len buffer under a static causal mask, matching the target's cache and
keeping the path export-friendly. The stateless is_causal forward is unchanged.

forward_cached requires batch size 1 and contiguous-from-0 writes (overwrites
for speculative rollback are allowed, gapped seeds are not); the invariant is
enforced by an eager-only validator that is skipped under export. Tests cover
cached prefill plus single-step decode against the stateless recompute, rollback
reseeding, and rejection of gapped/offset seeds and B>1.

Authored with assistance from Claude Code.

[digantdesai]

Stack from ghstack (oldest at bottom):

• Add EAGLE-3 end-to-end speculative-decode test #20157
• Add the EAGLE-3 speculative-decoding runner (CUDA) #20156
• Add the EAGLE-3 speculator CUDA export #20155
• Add a weight-stationary INT4 GEMM for small M #20154
• Add the Eagle3Speculator module for speculative decoding #20153
• -> Add KV cache to the EAGLE-3 draft head #20152
• Add Gemma4 EAGLE-3 eager reference #20151
• Add Gemma4 EAGLE-3 hidden-state taps #20150
• Add EAGLE-3 draft head #20149

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20152

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

⏳ No Failures, 3 Pending

As of commit d782162 with merge base dc55469 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[metascroy]

What distinguishes this from our standard KVCache? Can we just import the standard one?