Allow optarch values to be partial maps including vector extensions

[Flamefire]

Instead of only setting optimal compiler arguments based on architecture and CPU family/vendor also include the supported vector extensions as criteria to choose flags.
To simplify specifying generic flags allow partial matches, e.g. a fallback setting for any x86 arch which doesn't have more specific values.
Example:

    COMPILER_OPTIMAL_ARCHITECTURE_OPTION = {
        (systemtools.X86_64, ): 'xHost',
        (systemtools.X86_64, systemtools.AMD, systemtools.AVX2): 'mavx2',
    }

Closes #3793

[Flamefire]

@boegel Updated this PR

[Flamefire]

Ping on this as this came up in the EUM with respect to intel compilers

[joeydumont]

To add to this, it seems like spack has already done some work in mapping CPU feature flags to compiler versions, see microarchitectures.json. There is some logic in portage as well, where CPU_FLAGS variable map to USE flags, but this probably maps less closely to the problem addressed in this PR than what archspec does.

[Flamefire]

We just got hit by this not being implemented:

tf_gen.F90(172): catastrophic error: Function return parameter requires SSE register while SSE is disabled.

When compiling HDF5 with the Intel compiler and -xhost on AMD EPYC 7702