Add local Modal leaderboard mode

[msaroufim]

Summary

• add popcorn submit --local for running public evaluations in the user Modal workspace
• reuse KernelBot task construction and evaluation for test, benchmark, and leaderboard modes
• format per-case results and compute the configured leaderboard score locally
• bypass Popcorn registration and the hosted API for local runs
• document Modal token setup, supported GPUs, and unofficial-result limitations

Main-branch freshness

• resolve gpu-mode/reference-kernels and gpu-mode/kernelbot main to immutable commit SHAs on every invocation
• include those SHAs in the Modal image definition so a new upstream commit invalidates the cached image
• clone and evaluate those exact commits, then print both SHAs in the result
• retry GitHub resolution and fail closed instead of falling back to a potentially stale symbolic main
• allow explicit ref pinning with POPCORN_REFERENCE_KERNELS_REF and POPCORN_KERNELBOT_REF

Limitations

Local leaderboard mode reproduces the public test, benchmark, and ranked stages. It cannot reproduce server-only private seeds or create an official gpumode.com submission.

Test plan

• cargo fmt --all -- --check
• cargo test (53 passed)
• cargo clippy --all-targets --all-features -- -D warnings
• embedded Python helper syntax check
• git diff --check

A paid Modal GPU end-to-end run was not performed because this machine does not have Modal credentials configured.