Convert x86/sse41.rs intrinsics to const generics

[lqd]

This PR updates the x86 SSE 4.1 intrinsics to use const generics.

Of note, some of these made an effort to test the masking/clamping of the constify immediates macros, testing with both in-range and out-of-range values in:

• _mm_extract_ps: _mm_extract_ps(a, 5) on a 2-bit immediate
• _mm_extract_epi8: _mm_extract_epi8(a, 19) on a 4-bit immediate
• _mm_extract_epi32: _mm_extract_epi32(a, 5) on a 2-bit immediate
• _mm_insert_epi8: _mm_insert_epi8(a, 32, 17) on a 4-bit immediate
• _mm_insert_epi32: _mm_insert_epi32(a, 32, 5) on a 2-bit immediate

As discussed with Amanieu on zulip, I've made such out-of-range immediates errors with static assertions, and changed the tests in these cases. Opinions are welcome.

I've noticed our previous PRs were squashed when merged, so I didn't make a commit per intrinsic this time. Of course, I can do so if it would be helpful for the review ?

[rust-highfive]

r? @Amanieu

(rust-highfive has picked a reviewer for you, use r? to override)

[Amanieu]

It might be worth making this generic over the bit width:

struct ValidateConstImm8<const imm i32, const bits: i32>

The check would then be:

let _ = 1 / ((imm >= 0 && imm < (1 << bits)) as usize);

Then we can simply define macros referring to a single instance of this struct. These macros should ideally be in the generic macros.rs instead of the arch-specific one.

[lqd]

good idea, that should be enough for most of the other constify_imms analogues. (the other design has the advantage of encoding the immediate width in its name, and the expression shows up in the const error message, but a single struct will be better)