[Bugfix] Disable FlashInfer MLA prefill by default due to chunked prefill issues

[mgoin]

Purpose

Address #26042 for now to unblock the release. We will take a performance regression on B200 due to falling back to FA2 prefill, but we need correctness first.

Test Plan

Test Result

Base commands:
vllm serve deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
python tests/evals/gsm8k/gsm8k_eval.py

# Hopper

vllm serve deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
Accuracy: 0.794

# Blackwell

## Default backends (FlashInfer prefill and CUTLASS MLA decode)
vllm serve deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
Accuracy: 0.216

## Using FA2 prefill and CUTLASS MLA decode
VLLM_DISABLE_FLASHINFER_PREFILL=1 vllm serve deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
Accuracy: 0.785

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request updates the default value for VLLM_DISABLE_FLASHINFER_PREFILL to True, effectively disabling FlashInfer MLA prefill by default to address issues with chunked prefill. The change is applied consistently to both the type-hinted default and the environment variable parsing logic. While the change itself is correct, I've identified a related potential issue: this environment variable, which likely influences the computation graph, is not included in the cache key computation. This could lead to incorrect cache hits if the flag is toggled.

[pavanimajety]

Could we instead try to set this to "FA2"
https://github.com/vllm-project/vllm/blob/main/vllm/v1/attention/backends/mla/common.py#L573

FI pointer: https://github.com/flashinfer-ai/flashinfer/blob/main/flashinfer/prefill.py#L2397-L2401

[mgoin]

Unfortunately the FA2 backend in flashinfer still has the same issue, so it is likely vLLM preparing inputs improperly for flashinfer

[LucasWilkinson]

use: #26063 instead