Fix panics on out-of-range CurveID and oversized hash-to-curve domains

[adecaro]

Summary

Fixes the panics described in #51:

• Critical: Zr/G1/G2/Gt UnmarshalJSON indexed the Curves slice with an unbounded, attacker-controlled CurveID from the JSON payload, causing an unrecovered index out of range panic on malformed input (e.g. {"curve": 9999, ...}). Added a validCurveID bounds check that returns an error instead.
• Critical: Curve.HashToG1WithDomain/HashToG2WithDomain panicked on the gnark-backed curves when given a domain longer than 255 bytes. Wrapped both in recover(), consistent with the existing panic-to-nil pattern used elsewhere in math.go, returning nil instead.
• Low: Clarified FP256BN's HashToG2/HashToG2WithDomain panic message (AMCL has no G2 hash map) in both driver/amcl/fp256bn.go and fp256bn_miracl.go.
• Doc-only: Corrected Mul2/JointScalarMultiplication doc comments that implied a wall-clock speedup — verified against gnark-crypto source that neither implementation uses the GLV endomorphism, so the joint scalar-mult is an allocation optimization only, not a wall-clock one.
• Doc-only: Documented the rawBigInt dual-representation invariant on BLS12-381's Zr type (driver/gurvy/bls12381/bls12-381.go) so a future change doesn't silently drop GroupOrder's raw value via Plus/Minus/Mul.

Two AMCL-side performance findings (naive MultiScalarMul, Neg() via full scalar-mult instead of native negation) are intentionally out of scope for this PR — tracked as follow-up perf work, not correctness fixes.

Test plan

• Added TestJSONUnmarshalerInvalidCurveID — regression for the CurveID panic (rejects out-of-range/negative IDs with an error, still accepts all valid IDs).
• Added TestHashToGWithDomainOverlongDomain — regression for the oversized-domain panic across all curves, including AMCL's differing behavior (HMAC-based HashToG1WithDomain never panics; HashToG2WithDomain panics unconditionally, now caught by recover() and returned as nil).
• Added FuzzUnmarshalJSONCurveID and FuzzHashToGWithDomain native Go fuzz targets — ran locally for 20s each (~2.2M and ~209k executions) with zero crashes.
• go build ./..., go vet ./..., go test ./... all pass.

Closes #51.