Prebuild the benchmark binary with -O3 on Android

[lgeiger]

What do these changes do?

This PR makes sure that the Android benchmarking binary get's built with all optimizations enabled. Previously we also set -O3 in the Raspberry PI configs, I moved this now the the release built in the hope to slightly speedup ARM CI.

Benchmark Results

I didn't see any measurable difference on my system, but the generated binary has a slightly different size, so it is definitely doing something.

[lgeiger]

Looks like this change enabled debug checks for the ARM unit tests which now fail at:

compute-engine/larq_compute_engine/core/bgemm/ruy_pack.h

Lines 48 to 50 in fb25c8f

	// Ruy supports collecting column sums during the packing process, which we
	// have no need for.
	RUY_DCHECK_EQ(packed_matrix->sums, nullptr);

I think #529 touches the same code, I can try and debug what's going on after that has been merged.

[AdamHillier]

Ah yes I had long suspected that debug checks were for some reason disabled, and this confirms it haha. That specific check by the way was added by me and I'm pretty sure it can be removed.

[lgeiger]

That specific check by the way was added by me and I'm pretty sure it can be removed.

I'm fine with removing the check.

The matrices are prepared in:

compute-engine/larq_compute_engine/core/bgemm/bgemm.h

Line 80 in 590f8ef

ruy::PreparePackedMatrices(ruy_ctx, &bgemm_trmul_params);

Which sets the sums here. Do you think we should remove this instead, or would this break things?

[AdamHillier]

Do you think we should remove this instead, or would this break things?

I'm not sure what the impact of removing that call would be, so for now I would probably just remove the debug check.