feat(s2n-quic-core): add IP checksum function

[camshaft]

Description of changes:

When constructing raw IP and UDP packets, we need a way to compute the checksum for the appropriate fields. This PR implements the internet checksum function, as described in RFC1071.

I wanted to see if using SIMD/x86 intrinsics could improve performance at all over the generic version. Turns out that it's pretty significant! Below are the comparisons to Fuchsia's internet-checksum crate, which itself is pretty optimized.

inet/s2n/checksum/1500  time:   [27.122 ns 27.129 ns 27.136 ns]
                        thrpt:  [51.482 GiB/s 51.494 GiB/s 51.508 GiB/s]
inet/fuchsia/checksum/1500
                        time:   [59.451 ns 59.478 ns 59.504 ns]
                        thrpt:  [23.477 GiB/s 23.488 GiB/s 23.498 GiB/s]
inet/s2n/checksum/9000  time:   [146.39 ns 146.41 ns 146.44 ns]
                        thrpt:  [57.238 GiB/s 57.249 GiB/s 57.258 GiB/s]
inet/fuchsia/checksum/9000
                        time:   [359.83 ns 359.86 ns 359.89 ns]
                        thrpt:  [23.290 GiB/s 23.292 GiB/s 23.294 GiB/s]
inet/s2n/checksum/65536 time:   [1.0878 µs 1.0879 µs 1.0882 µs]
                        thrpt:  [56.091 GiB/s 56.102 GiB/s 56.110 GiB/s]
inet/fuchsia/checksum/65536
                        time:   [2.3346 µs 2.3348 µs 2.3349 µs]
                        thrpt:  [26.140 GiB/s 26.142 GiB/s 26.143 GiB/s]

Call-outs:

I tried adding a differential test using the internet-checksum crate, but it panics on integer overflow. This has been fixed in the repository but has not been published to crates.io.

Testing:

• Differential test with the RFC implementation
  • Runs under kani and miri
• Differential test to compare the x86 implementation to the generic implementation

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[WesleyRosenblum]

I think we'd always be above this with the min_indistinguishable_packet_len being enforced. Though I'm fine keeping this here as well

[camshaft]

So that's for the whole payload, yes. But smaller values (like just the IPv4 pseudo headers) can also be written.

[WesleyRosenblum]

Suggested change

	// Since we have 16 bytes as input, we need to outputs since we're doubling the bit-width
	// Since we have 16 bytes as input, we need two outputs since we're doubling the bit-width