Add Gemma4 EAGLE-3 hidden-state taps

[digantdesai]

Adds validated tap-layer configuration, a shared decode path that can
return concatenated tap features, and forward_logits_taps for
soft-capped logits with EAGLE-3 features. Includes tests for tap
ordering, validation, logits parity, and last-logits behavior.

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/20150

• 📄 Preview Python docs built from this PR

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

⏳ No Failures, 1 Pending

As of commit 57ef8a5 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]

For tapped model, do we need all logits, not just final so sampling can be done outside model?

If so, why is last_logits_only default to True on forward_logits_taps?