Improve circuit setup phase

[frozenspider]

Resolves #45, also partially implements #44

This PR makes R1CS circuit setup faster, use less peak memory and do significantly less allocation:
(Using circuit_setup benchmark for measurement)

sha256-2kb	Before	After
Time	~700 ms	~300 ms
Max alloc	260 MB	140 MB
Total alloc	2505 MB	372 MB

Summary of changes:

• Tweaked bigint_to_fq to use LazyLock to only map Q100 ($2^{100} - 15$) to BigInt once instead of on every invocation. Part in matrix setup: 92% -> 88%.
• Parallelized SparseMatrix::map_values_with.
• Replaced BTreeMap/BTreeSet in LinearCombination and BoolLinearCombination with Vec that are kept sorted and deduplicated.
• Improved efficiency of ConstraintGenerator::bitz_unsigned.
• Additionally, map_values_with now takes a function by value instead of function reference - this is a style change, performance is unaffected.

Sampling with:

$ CARGO_PROFILE_BENCH_DEBUG=line-tables-only RUSTFLAGS="-C target-cpu=native -C force-frame-pointers=yes" cargo bench -p bitz-cli --bench circuits --no-run
$ samply record -- target/release/deps/circuits-<HASH> --bench circuit_setup --sample-count 20 --sample-size 1

Final measurements done with:

RUSTFLAGS="-C target-cpu=native -C force-frame-pointers=yes" cargo bench -p bitz-cli --bench circuits -- circuit_setup --sample-count 20 --sample-size 1

[zkfriendly]

LGTM, I was able to reproduce the results

[zkfriendly]

Noticed the CI is failing; can we address that here before merging?

[frozenspider]

@zkfriendly: this is a separate issue with our test suite being quite borked (see #63, but it's not just that).
Addressing it might not be trivial, so I'd prefer to do it separately.