ggml webgpu: profiling, CI updates, reworking of command submission

[reeselevine]

This PR adds:

• Optional profiling for the host (CPU) and device (GPU) sides, which will be helpful when working on optimization.
• Bumps WebGPU Ubuntu CI testing to use 24.04 to match Vulkan
• Reworks command submission to avoid global locks on pushing to staging vectors, and instead have each thread submit work in graph_compute.
• As part of that work, I was running into issues where deadlock was occurring, either because the parameter buffer pool was exhausted, or in some cases due to what I believe are underlying issues in Dawn's thread-safety approach. Serializing each command submission seems to avoid issues, so added options for that, and left a TODO in the code to explore other solutions.
• For now, I'm comfortable leaving it as is, because one of the next major goals is getting the WebGPU backend running fully in the browser, which will likely be in a single-threaded context to start, as mentioned in webgpu : fix build on emscripten #15826.

[CISC]

BTW, consider adding yourself to CODEOWNERS.

[CISC]

@reeselevine I think you have the ability, merge at will.

[reeselevine]

@CISC yeah I will, I was just at the gym this morning and thought of a potential solution for my deadlock problems, so trying a couple commits to fix that, then I will merge!

[reeselevine]

@CISC looks like the inflight threads checks fixed the deadlock issues! Thanks for your review of the logic. I'm going to remove the SERIALIZE option and then I'll merge.