precompile in the current process

[shunting314]

Stacked PRs:

• add kernel-filter to select kernel for allreduce-rmsnorm #1744
• APIs to debug distributed kernel #1743
• helion distributed kernel autotuning #1532
• ->precompile in the current process #1730

---

precompile in the current process for distributed kernels.

    # The reason we need this is due to some tricky distributed kernels
    # like https://gist.github.com/shunting314/81f13ce00f835b21ab6466e21454b7c5 . We specialize the RANK argument for each GPU,
    # some rank may get out of resouce errors while others don't
    # due to the specialization.
    #
    # Without precompilation here, some rank may fail and skip running
    # the kernel while outer ranks waiting for its peers. It
    # results in a stuck job.
    #
    # Precompiilation happening in child process is not enough because
    # CUDA is not available there. We can not check resource usage
    # like shared-memory, tmem, max-threads etc.
    #
    # This precompilation has overhead. Only do it if distributed is
    # initialized.

Test is done in the next PR

[shunting314]

test failures unrelated. The timeout one is not related since either autotuning is not enabled in the test or distributed is not initialized (thus the PR is a nop)