RUST-2239 Fix performance of bson 3.0

[abr-egn]

RUST-2239

This takes 3.0 benchmark performance from a 0-90% regression, depending on the benchmark, to a 0-50% improvement 😁

All of the delta is in converting a Document into bytes; the improvement is from (a) encoding everything into a single running buffer rather than letting subdocuments allocate their own and (b) avoiding copies when encoding individual elements.

Benchmark results from 2.15.x:

TEST: Flat BSON Decoding – Score: 299.481 MB/s, Median Iteration Time: 0.251s
TEST: Flat BSON Encoding – Score: 2204.125 MB/s, Median Iteration Time: 0.034s
TEST: Deep BSON Decoding – Score: 71.221 MB/s, Median Iteration Time: 0.276s
TEST: Deep BSON Encoding – Score: 580.878 MB/s, Median Iteration Time: 0.034s
TEST: Full BSON Decoding – Score: 212.965 MB/s, Median Iteration Time: 0.269s
TEST: Full BSON Encoding – Score: 1189.535 MB/s, Median Iteration Time: 0.048s
BSONBench Score = 759.701 MB/s

Benchmark results from 3.0 before this PR:

TEST: Flat BSON Decoding – Score: 292.714 MB/s, Median Iteration Time: 0.257s
TEST: Flat BSON Encoding – Score: 1435.658 MB/s, Median Iteration Time: 0.052s
TEST: Deep BSON Decoding – Score: 74.274 MB/s, Median Iteration Time: 0.264s
TEST: Deep BSON Encoding – Score: 62.359 MB/s, Median Iteration Time: 0.315s
TEST: Full BSON Decoding – Score: 244.282 MB/s, Median Iteration Time: 0.235s
TEST: Full BSON Encoding – Score: 477.151 MB/s, Median Iteration Time: 0.120s
BSONBench Score = 431.073 MB/s

Benchmark results from 3.0 after this PR:

TEST: Flat BSON Decoding -- Score: 357.983 MB/s, Median Iteration Time: 0.210s
TEST: Flat BSON Encoding -- Score: 2447.073 MB/s, Median Iteration Time: 0.031s
TEST: Deep BSON Decoding -- Score: 73.080 MB/s, Median Iteration Time: 0.269s
TEST: Deep BSON Encoding -- Score: 910.454 MB/s, Median Iteration Time: 0.022s
TEST: Full BSON Decoding -- Score: 278.466 MB/s, Median Iteration Time: 0.206s
TEST: Full BSON Encoding -- Score: 1573.123 MB/s, Median Iteration Time: 0.036s
BSONBench Score = 940.030 MB/s

[abr-egn]

This is the "avoiding copies on individual elements" part; for the most part this can just delegate to RawBsonRef. Unfortunately, for the more complex types that aren't just a simple reference conversion there's some duplication of logic that was unavoidable. I reworked the RawWriter helper into DocWriter to boil down what I could.

[abr-egn]

With Bson::append_to now present, the "share a buffer while encoding" becomes very straightforward.

[isabelatkinson]

nice!