optimize the fixed32 and fixed64 encoding

[tardyp]

The current implementation of write_raw_little_endian32 is just using self.write_raw_bytes.

As this go through the function boundaries, the compiler is not smart enough to simplify the data copy, and end-up calling memcpy.

This inlined implementation is twice as fast as the previous according to the attached bench.

test bench_write_raw_fixed32   ... bench:       5,860.52 ns/iter (+/- 62.51)
test bench_write_raw_fixed64   ... bench:       4,303.08 ns/iter (+/- 29.47)
test bench_write_raw_varint_32 ... bench:       2,555.20 ns/iter (+/- 126.98)

now fixed32/64 is as expected a bit faster than varint

test bench_write_raw_fixed32   ... bench:       2,143.48 ns/iter (+/- 21.70)
test bench_write_raw_fixed64   ... bench:       2,209.28 ns/iter (+/- 32.42)
test bench_write_raw_varint_32 ... bench:       2,480.89 ns/iter (+/- 88.95)

[tardyp]

I think that the CI failure is a flaky test. let me know if you disagree.