GDN Decode kernel optimization

[helloworld1]

Description

Optimize gdn decode kernel

• Fuse silu in decode kernel, this is almost free to do in kernel
• For GQA, compute qk first and then repeat instead of repeat qk and then compute.
• Default block size is not optimal for pipelining. We have enough VMEM to fit everything in 1 block but we want more blocks to overlap the DMAs in first / last block

+--------+---------+---------+--------------+------------+------------+--------------+
| Decode | GDN Old | GDN New | GDN % | Kernel Old | Kernel New | Kernel % |
| Tokens | (us) | (us) | Difference | (us) | (us) | Difference |
+--------+---------+---------+--------------+------------+------------+--------------+
| 64 | 88.9 | 61.3 | -31.05% | 53.6 | 36.7 | -31.53% |
| 128 | 130.8 | 93.7 | -28.36% | 93.6 | 66.6 | -28.85% |
| 256 | 218.8 | 164 | -25.05% | 173.4 | 126 | -27.34% |
| 512 | 377 | 283 | -24.93% | 332 | 239 | -28.01% |
+--------+---------+---------+--------------+------------+------------+--------------+

Tests

Tested E2E benchmark.
Qwen 3.5 397B, total token throughput
1K/8K c64: 2051 -> 2163 (+5.5%)
8K/1K c64: 10745 -> 11522 (+7.2%)

Tested lm_eval mmlu_pro no regression

|       Tasks       |Version|    Filter    |n-shot|  Metric   |   |Value |   |Stderr|
|-------------------|------:|--------------|-----:|-----------|---|-----:|---|-----:|
|mmlu_pro           |    2.0|custom-extract|     5|exact_match|↑  |0.8286|±  |0.0139|
| - biology         |    3.1|custom-extract|     5|exact_match|↑  |0.9400|±  |0.0339|
| - business        |    3.1|custom-extract|     5|exact_match|↑  |0.8800|±  |0.0464|
| - chemistry       |    3.1|custom-extract|     5|exact_match|↑  |0.9400|±  |0.0339|
| - computer_science|    3.1|custom-extract|     5|exact_match|↑  |0.8400|±  |0.0524|
| - economics       |    3.1|custom-extract|     5|exact_match|↑  |0.9000|±  |0.0429|
| - engineering     |    3.1|custom-extract|     5|exact_match|↑  |0.6600|±  |0.0677|
| - health          |    3.1|custom-extract|     5|exact_match|↑  |0.7800|±  |0.0592|
| - history         |    3.1|custom-extract|     5|exact_match|↑  |0.6800|±  |0.0666|
| - law             |    3.1|custom-extract|     5|exact_match|↑  |0.7400|±  |0.0627|
| - math            |    3.1|custom-extract|     5|exact_match|↑  |0.9600|±  |0.0280|
| - other           |    3.1|custom-extract|     5|exact_match|↑  |0.7400|±  |0.0627|
| - philosophy      |    3.1|custom-extract|     5|exact_match|↑  |0.8200|±  |0.0549|
| - physics         |    3.1|custom-extract|     5|exact_match|↑  |0.9400|±  |0.0339|
| - psychology      |    3.1|custom-extract|     5|exact_match|↑  |0.7800|±  |0.0592|

| Groups |Version|    Filter    |n-shot|  Metric   |   |Value |   |Stderr|
|--------|------:|--------------|-----:|-----------|---|-----:|---|-----:|
|mmlu_pro|      2|custom-extract|     5|exact_match|↑  |0.8286|±  |0.0139|

Checklist

Before submitting this PR, please make sure:

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have made or will make corresponding changes to any relevant documentation.

[github-actions]

Description

Start with a short description of what the PR does and how this is a change from
the past.

The rest of the description includes relevant details and context, examples:

• why is this change being made,
• the problem being solved and any relevant context,
• why this is a good solution,
• some information about the specific implementation,
• shortcomings of the solution and possible future improvements.

If the change fixes a Github issue, please include a link, e.g.,:
FIXES: #123456

Tests

Please describe how you tested this change, and include any instructions and/or
commands to reproduce.

Checklist

Before submitting this PR, please make sure:

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have made or will make corresponding changes to any relevant documentation.