Eliminate temporary buffer allocations during bson serialization

[Edarke]

Fixes: JAVA-5786

• Override Outputbuffer.toByteArray() to avoid extraneous array copy
• Modify BsonBinaryWriter to call OutputBuffers.writeObjectId so implementations can opt to use putToByteBuffer rather than allocating a temporary byte[]
• Wrap array in ByteBuffer so we can write numeric values 8 bytes at a time

Motivation

Atlas Search uses the Java driver to communicate with the Mongodb Server. We serialize up to ~300k documents per batch, primarily consisting of ObjectID and numbers. For these large batches, serialization has measurable overhead. This PR proposes some simple optimizations for BasicOutputBuffer without altering the API.

JMH Benchmarks

Benchmark	Main (ops/ms)	GC (B/op)	This PR (ops/ms)	GC (B/op)
toByteArray_large	0.907	26,777,857	2.269	13,388,928
toByteArray_small	110,515	160	184,816	80
encodeId	23,774	134	32,696	104
write_id	28,255	168	36,467	80
write_numbers	16,327	80	18,125	80

[jyemin]

@Edarke please open a Jira issue in the JAVA project for this PR, as per our contribution guide.

Thanks!

[jyemin]

It seems error-prone (and confusing) to add this field without completely replacing use of the two existing fields with the new one. For instance, you have to always remember to use absolute position when writing to buf, so that its own position is ignored in favor of position field.

What do you think about getting rid of buffer and position fields entirely?

[Edarke]

Agreed. Updated in 0385da6