Force hashing to be done on aligned usizes

[krtab]

Summary

This is a WIP/PoC of doing hashing on aligned usizes. This seems to reduce the size of the generated assembly, and I would like to benchmark it, but I don't know how to run a compiler bench using a modified crate.

Drawbacks of this PR:

• Adds unsafe code
• Changes the generated hashes

Explanation

Let's call N the size in bytes of a usize.
Two cases are distinguished. First if bytes.len() < N, then the previous strategy is used, that is 32 bits are read, then 16, then 8.

However, if the slice is at least N long, we can:

1. Extract the first N bytes in a usize (which may not be aligned),
2. Extract the last N bytes in a different usize (which may not be aligned),
3. Extract a correctly aligned usize slice. On real code (ie not miri), align_to guarantees that this slice is maximal

These three elements may overlap, but they cover all input bytes. We can now add to the hasher all these usizes, and the compiler will take advantage of the middle slice being well aligned.

Remarks

I haven't updated the tests so it will inevitably fail CI.

[krtab]

I've found how to do the benchmarking here: rust-lang/rust#59594

[Noratrieb]

Yes, make a PR to rust-lang/rust changing the dependency to your branch.

[krtab]

I've realized that my algorithm is buggy. Hashing two equal byte slices will lead to different results if their start address is not the same modulo 64.