Fix: fix the deterministic issue in the MTP Eagle path

[lfr-0531]

Description

This PR updates the first draft forward in the MTP Eagle path to use all accepted tokens instead of just the last one. This allows the KV cache for the draft layer to be updated.

Before this fix, each iteration, the KV cache didn't store the key/value pair for the last draft token (let's call it D) because it was the output of the last draft forward pass. For this last draft token, we call it D. In the next iteration, if all draft tokens were accepted, we'd use the newly generated tokens as inputs for the first draft forward. But since the KV cache had incorrect key/value data for token D, identical inputs could produce different draft tokens.

With this fix, the KV cache will be updated in the first draft forward each iteration, making MTP Eagle deterministic.

I tested the DS-R1-FP4 model with the same dataset and on the same node (with BS=1). The acceptance rate will increase a little bit:

Before the changes:

===========================================================
= PERFORMANCE OVERVIEW 
===========================================================
Request Throughput (req/sec):                     0.1530
Total Output Throughput (tokens/sec):             313.0214
Total Token Throughput (tokens/sec):              472.7368
Total Latency (ms):                               65372.5274
Average request latency (ms):                     6537.2043
Per User Output Throughput [w/ ctx] (tps/user):   313.3934
Per GPU Output Throughput (tps/gpu):              39.1277

-- Acceptance Rate Details --------------------------------
[AR] MINIMUM: 2.69
[AR] MAXIMUM: 3.08
[AR] AVERAGE: 2.81
[AR] P50    : 2.80
[AR] P90    : 3.08
[AR] P95    : 3.08
[AR] P99    : 3.08

After:

===========================================================
= PERFORMANCE OVERVIEW 
===========================================================
Request Throughput (req/sec):                     0.1544
Total Output Throughput (tokens/sec):             316.0127
Total Token Throughput (tokens/sec):              477.2701
Total Latency (ms):                               64747.4007
Average request latency (ms):                     6474.6948
Per User Output Throughput [w/ ctx] (tps/user):   316.3113
Per GPU Output Throughput (tps/gpu):              39.5016

-- Acceptance Rate Details --------------------------------
[AR] MINIMUM: 2.75
[AR] MAXIMUM: 3.07
[AR] AVERAGE: 2.83
[AR] P50    : 2.84
[AR] P90    : 3.07
[AR] P95    : 3.07
[AR] P99    : 3.07

For the model accuracy, with this fix, I enabled MTP Ealge and tested GPQA diamond twice with the same random seed. We got the same results 70.202 ± 3.2586, which is expected.

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #9204 [ run ] triggered by Bot

[lfr-0531]

/bot run

[lfr-0531]

/bot kill

[tensorrt-cicd]

PR_Github #9264 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9265 [ kill ] triggered by Bot

[tensorrt-cicd]

PR_Github #9264 [ run ] completed with state ABORTED

[tensorrt-cicd]

PR_Github #9265 [ kill ] completed with state SUCCESS
Successfully killed previous jobs for commit 9fcb06c

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #9277 [ run ] triggered by Bot

[lfr-0531]

/bot kill

[tensorrt-cicd]

PR_Github #9307 [ kill ] triggered by Bot

[tensorrt-cicd]

PR_Github #9277 [ run ] completed with state ABORTED

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #9307 [ kill ] completed with state SUCCESS
Successfully killed previous jobs for commit 0bf3101

[tensorrt-cicd]

PR_Github #9308 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9308 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #6830 completed with status: 'FAILURE'

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #9334 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9334 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #6851 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.