bench(exla): time matmul compile separately from warm runs

[Chapaman]

mix run bench/matmul.exs times an f32 matmul at 64, 256, and 1024 on the host client. Cold, shape-cache miss, and shape-cache hit are single calls. Warm runs go through Benchee after that shape has already been compiled. Inputs are on device before the timer starts, and the executable waits for the result buffer, so a sample is not just a launch.

On this machine the first shape compiled in 334 ms. The next two compiled in 18 ms and 12 ms, and each emitted one [:exla, :compilation] event. Warm runs and the hits emitted none. Steady state was about 79 μs, 1.6 ms, and 27 ms. Nothing in that suggested a cache change.

[polvalente]

I think we could turn this into an example livebook instead?

Seems like a good opportunity to teach about compiler cache hits and misses, and how that can affect benchmarks.

Also, Benchee.run has a warm time option, so any compilation times can be separated out using that instead of forcing compilation before running

[Chapaman]

Moved it to guides/compilation-cache.livemd. The first cells time a miss, a hit, and a new shape. Benchee uses warmup for 1024, so that compile stays out of the samples without a call beforehand.