Zstandard decompression performance

[dweiller]

The Zstandard decompression implementation has not had any performance benchmarking and optimisation work done.

To make sure we have a speedy implementation we should:

• include a decompression benchmark utility in lib/std/compress/zstandard/, and/or add a benchmark to gotta-go-fast
• compare performance with the reference implementation - unless it requires more implementation complexity or generated code size than is desired, there is no reason we can't be as fast or faster than the reference.

Extracted from #14702.

One thing to think about when optimising is that the current implementation does not decode 4-stream literal sections in parallel in order to avoid memory allocations (or remove a need for an additional 128KiB of stack memory) in the non-allocating API. This presumably sacrifices decompression performance, but implementing and benchmarking a parallel implementation is needed to assess what the impact is in practice. An implementation that decodes literal sections in parallel would require re-working the block decoding in lib/std/compress/zstandard/decode/block.zig to eagerly decode literals into an extra internal buffer using four threads. Another (probably better) option is to decode the streams 'in parallel' on a single core by exploiting instruction level parallelism on super-scalar processors.