Use memcmp for field element comparisons

[jtraglia]

In compute_cells_and_kzg_proofs, comparing field element limbs directly is 2.87% faster with precompute=8.

Note: I benchmarked verify_cell_kzg_proof_batch too but there wasn't any difference.

Before:

$ go clean -cache && GOMAXPROCS=1 go test -test.bench=Benchmark/ComputeCellsAndKZGProofs -test.run=^$ ./bindings/go 
goos: darwin
goarch: arm64
pkg: github.com/ethereum/c-kzg-4844/v2/bindings/go
cpu: Apple M1
Benchmark/ComputeCellsAndKZGProofs(precompute=0)                       4         263398458 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=1)                       2         908959958 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=2)                       3         488505694 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=3)                       3         350817750 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=4)                       4         277039625 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=5)                       5         240792967 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=6)                       5         213914367 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=7)                       6         198406333 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=8)                       6         184162486 ns/op
Benchmark/ComputeCellsAndKZGProofsParallel(count=8)                    4         289445594 ns/op
PASS
ok      github.com/ethereum/c-kzg-4844/v2/bindings/go   96.002s

After:

$ go clean -cache && GOMAXPROCS=1 go test -test.bench=Benchmark/ComputeCellsAndKZGProofs -test.run=^$ ./bindings/go
goos: darwin
goarch: arm64
pkg: github.com/ethereum/c-kzg-4844/v2/bindings/go
cpu: Apple M1
Benchmark/ComputeCellsAndKZGProofs(precompute=0)                       4         263577667 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=1)                       2         904966354 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=2)                       3         482739042 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=3)                       3         345744278 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=4)                       4         276764583 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=5)                       5         238037950 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=6)                       5         211482933 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=7)                       6         195122521 ns/op
Benchmark/ComputeCellsAndKZGProofs(precompute=8)                       6         178871958 ns/op
Benchmark/ComputeCellsAndKZGProofsParallel(count=8)                    4         290647542 ns/op
PASS
ok      github.com/ethereum/c-kzg-4844/v2/bindings/go   95.532s

[b-wagn]

LGTM!

[asn-d6]

This switch to memcmp is not truly straightforward because the old approach (using blst_uint64_from_fr()) was doing a modular reduction on the input field element. Whereas now we don't, and this could be a problem if this function ever receives an unreduced field element.

I don't think this is a problem, because we should not be handling unreduced field elements, but it's a new assumption that should be taken into account every time we parse or introduce field elements in any way.

[asn-d6]

Hmm, FR_NULL is only used during recovery and nowhere else. Perhaps we can just kill FR_NULL completely, and initialize recovered_cells_fr with FR_ZERO. After all, the missing cells will be set to zero by multiplication with vanishing_poly_eval. This is what's done in the spec right now (see extended_evaluation_rbo).

[jtraglia]

Oh you're right. I like this suggestion. Let me do this.

[kevaundray]

Can these be removed?

[jtraglia]

Oops good catch, removed.

[kevaundray]

Maybe we should write something like "field elements are assumed to be reduced" ?