[RELEASE ONLY CHANGES] Pin test-infra workflows to release/2.13

[JacobSzwejbka]

Release-only changes. Do not merge to main.

Pins PyTorch test-infra reusable workflows, actions, and explicit test-infra-ref inputs to release/2.13. It updates release documentation clone commands from viable/strict to release/1.4, following scripts/release/apply-release-changes.sh.

Also carries the 1.3 release full-history link-check fix. Shallow checkout left the base-to-head range unavailable on branch creation and PR merge refs; fetch-depth: 0 makes URL, xref, and file-size lint comparisons reliable.

Test plan:

• git diff --check
• Parsed all 65 workflow YAML files with the system YAML parser
• Verified all 18 distinct referenced actions/workflows exist at pytorch/test-infra@release/2.13
• Verified .github/workflows/_link_check.yml matches the final 1.3 release fix
• bash scripts/lint_file_size.sh origin/release/1.4 HEAD
• Reran all nine failed release/1.4 branch-cut workflows

Authored with Codex.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21401

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures, 1 Unrelated Failure, 1 Unclassified Failure

As of commit d09609f with merge base cd380e7 ():

NEW FAILURES - The following jobs have failed:

• cuda-perf / export-models (SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4, quantized-int4-tile-packed, SocialLoca... / linux-job (gh)
  RuntimeError: Command bash /__w/_temp/exec_script failed with exit code 1
• Test CUDA Builds / export-model-cuda-artifact (SocialLocalMobile, Qwen3.5-35B-A3B-HQQ-INT4, quantized-int4-tile-packed) / linux-job (gh)
  RuntimeError: Command bash /__w/_temp/exec_script failed with exit code 1

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

• nightly / test-cortex-m-e2e-nightly / run (ds_cnn, cortex-m55) / ds_cnn-cortex-m55 (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  RuntimeError: Command docker exec -t 23aa06ac8691b885c90c7617c8b1b28701b6faa8400fbeeb14ac5df57eb60c23 /exec failed with exit code 1

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• cuda-perf / benchmark-cuda (SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4, quantized-int4-tile-packed, SocialLoc... / linux-job (gh) (trunk failure)
  [OSDC] Step script exited with code 1. This is a script/workflow error, not an infrastructure issue. Check the step logs above for the actual failure.

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

[JacobSzwejbka]

1.3 release-only audit:

Already present in the 1.4 cut:

• macOS TorchInductor cache isolation
• Parakeet tokenizer C++20 fix
• Cortex-M direct dependency metadata cleanup (superseded by the current packaging layout)

Carried in this PR after reproducing the same failure:

• full-history checkout for URL, xref, and file-size lint comparisons

Still needed before the 1.4 RC, in follow-up release-only changes:

• set version.txt to 1.4.0 and update SwiftPM/Raspberry Pi version examples
• finalize released dependency metadata for torch 2.13, torchao 0.17, torchcodec 0.15, and pytorch-tokenizers 1.4 after that package is published; update the tokenizer submodule to the v1.4.0 tag
• carry the 120-minute iOS Device Farm timeout, dynamic Android golden-artifact discovery, and mutable-wheel no-cache behavior

The 1.3 selective workflow unpins and MLX unpin were reactive fixes for release/2.12 input-schema/startup failures. The current release/2.13 calls are starting successfully, so those exceptions are not copied unless this PR produces equivalent failures.

[JacobSzwejbka]

Cherry-picked f6eaba239 (Flakey CI fix (#21110)) from main. The release branch reproduced the same MLX chunked-prefill mismatch twice (0.0107117 absolute difference against atol=0.01); this carries the upstream cross-hardware tolerance of atol=0.03 while retaining the greedy-token correctness assertion. The cherry-picked patch ID matches upstream exactly.