perf(s2n-quic-core): optimize varint encoding

[camshaft]

Description of changes:

This PR focuses on improving varint encoding a bit by:

• improving the table function to be completely branchless (see https://godbolt.org/z/9xrxd1osd)
• Adding a special case when we have at least 8 bytes in the target buffer and just doing a single 64 bit mov, rather than having to branch based on our integer size:
  

  s2n-quic/quic/s2n-quic-core/src/varint/table.rs

  Lines 262 to 271 in b847852

  	// optimize for the case where we have at least 8 bytes left and just write a full u64 to
  	// the buffer, but incrementing the offset by `self.len`.
  	//
  	// ignore this optimization under miri, since we're technically reading beyond the slice
  	// that the encoder gives us, which miri complains about.
  	if encoder.remaining_capacity() >= 8 && !cfg!(miri) {
  	self.encode_oversized(encoder);
  	} else {
  	self.encode_maybe_undersized(encoder);
  	}

I've also cleaned the code up a bit by moving the table functions into a separate module with an named struct, rather than a (u64, usize, u64).

Performance has improved by about 30-40%

varint/encode/array/1   time:   [1.6209 ns 1.6211 ns 1.6212 ns]
                        change: [-40.046% -39.123% -37.678%] (p = 0.00 < 0.05)
varint/encode/slice/1   time:   [2.4321 ns 2.4329 ns 2.4341 ns]
                        change: [-33.393% -33.335% -33.266%] (p = 0.00 < 0.05)
varint/encode/array_16/1
                        time:   [38.505 ns 38.515 ns 38.529 ns]
                        change: [-31.424% -31.257% -31.005%] (p = 0.00 < 0.05)
varint/encode/slice_16/1
                        time:   [38.676 ns 38.790 ns 38.889 ns]
                        change: [-38.523% -38.430% -38.354%] (p = 0.00 < 0.05)

Testing:

The existing tests should prevent any regressions in behavior.

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

[WesleyRosenblum]

2000! 🚀

[WesleyRosenblum]

Can you summarize (maybe in the PR description) how the optimization works and why it achieves 30-40% better perf?