Perf: Improve compiler speed

[anudit]

While testing our circuits we found some performance improvements.
This speeds up compilation ~2x across the board.

Changes

#	Change	Where	Gain
1	execute_component deep-cloned the whole ComponentRepresentation (hashmap of memory slices, each cloning AssignmentState + AST Meta) on every signal access into a subcomponent. Now access_values_by_reference - read-only there anyway, same bounds checks and errors. This clone is what the 2.1.9→2.2.x component/bus rework introduced.	constraint_generation/src/execute.rs	keccak256 24→13 s
2	FxHash instead of SipHash on the simplification/constraint maps, they're keyed by signal ids and compiler-controlled strings, so DoS resistance buys nothing and cost ~13 % of simplification. No new dependency. write_constraint_usize and apply_correspondence are now generic over the hasher.	circom_algebra/src/fast_hash.rs (new); circom_algebra, constraint_list, dag	jwt1024 19.1→16.5 s
3	mimalloc as global allocator. The profile's ~20 % allocator self-time was the system allocator being slow	circom/src/main.rs	jwt1024 16.5→12.2 s
4	modulus(a,b) computed ((a % b) + b) % b; since a % b is already in (-b, b) for positive b, the second division was always a no-op.	circom_algebra/src/modular_arithmetic.rs	jwt1024 12.2→10.6 s
5	lto = "thin", codegen-units = 1, panic = "abort" - the workspace had no [profile.release] at all.	Cargo.toml	~5 %

Bug fixes

Two circom_algebra tests already failed on unmodified a100fae;

• complement didn't reduce its operand. It truncates to the field's bit width before flipping, so an operand >= field lost its high bits and ~~x != x mod p. Now reduces into [0, field) first, like every other op in the module.
• algebra_constraint_apply_substitution skipped a step. apply_substitution deliberately leaves the raw form (explicit constant coefficients, zero or not); its internal fix_constraint is commented out because the pipeline calls it afterwards. The test now does the same.

Benchmarks on some of our circuits

Apple M2 Max (12 cores, 64 GB, macOS arm64). hyperfine with warmup, heavy circuits
×3 runs, light ×10. Flags: --r1cs --wasm --c -l ./node_modules. All three binaries
measured identically.

Circuit	Constraints	v2.2.3 after	v2.2.3 before	v2.1.9	Improvement
keccak256	239,211	6.00 s	18.76 s	9.77 s	3.13×
jwt1024	685,509	7.87 s	15.63 s	16.61 s	1.99×
oauth512	683,083	7.35 s	14.05 s	14.91 s	1.91×
jwt512	420,292	4.48 s	8.86 s	10.21 s	1.98×
jwt256	287,555	2.92 s	5.83 s	6.70 s	2.00×
aadhaar	164,841	1.40 s	2.86 s	4.07 s	2.04×
rsa	153,137	0.99 s	1.95 s	2.84 s	1.98×
sha256	31,699	0.72 s	1.37 s	1.42 s	1.92×
semaphore20	15,875	0.49 s	0.78 s	1.14 s	1.58×
loginProof	5,244	0.57 s	0.85 s	1.11 s	1.49×
merkle	10,464	0.42 s	0.64 s	0.79 s	1.52×