feat: Add explicit simd implementation for one pass svd in large action spaces.

[zwd-ms]

Add an explicit SIMD implementation using AVX-512 for the one pass svd part in large actions. For now, it only works with quadratic interactions.

By default, VW builds with VW_BUILD_LAS_WITH_SIMD=ON so it includes the SIMD binaries. The compile flag can be turned off, for example when some CI runs on unsupported architectures (or when someone building on really old platforms), but all tests should build and pass on a recent CPU.

To enable vectorization at runtime, pass in --explicit_simd on vw command line. This flag is advisory: it only works if the program detects that the platform actually supports the needed instructions.

Next steps include adding alternative implementations using other extension sets (e.g. AVX2), and adding automatic switching between scalar and simd code paths using CPUID info at runtime, instead of using the command-line flag.

[olgavrou]

A separate CI job should be added that turns on the compile time flag and runs the related tests (like we had for LAS while it was behind a compile time flag). It can run only for linux architectures

[zwd-ms]

A separate CI job should be added that turns on the compile time flag and runs the related tests (like we had for LAS while it was behind a compile time flag). It can run only for linux architectures

Good idea. I'm adding a CI in a subsequent PR that handles the runtime detection of supported instructions. Before that's done, the tests in CI won't always pass due to "illegal instructions" when the CI runs on unsupported machines.
Update: after adding the cpu feature detection, the compile flag is set to ON by default, so we can use the existing CI for tests now.