Carry generic proving improvements onto the public backend interfaces:
parallel witness scheduling and lookup rows, cached hash preparation,
partitioned lowering, and invalid-challenge rejection. Preserve checked
circuit assignments and keep KZG-specific implementations in ix.

Add auxiliary-device host LDEs and mixed quotient parity coverage, and pin
the sppark fork containing the shared synchronization and profiling fixes.

Constraint-free circuits take the portable quotient path instead of
panicking in the device evaluator; flatten_to_base reserves its output
once instead of allocating per element; a host-LDE worker whose offload
panicked stays out of service instead of poisoning later proofs.

Validation: pinned Rust 1.98.1 formatting and CPU/CUDA Clippy pass with
warnings denied. Focused parallel and serial CPU tests, public API tests,
and ten CUDA checks pass. Two-device checks and full suites were not run.

`ByteValue` carried its nibble decomposition eagerly, so `select` paid for
the value and both nibbles: 9.5 gates per byte, with the 64 byte selects
that order a Merkle pair costing 608 of a level's 672 gates. The compact
BLAKE3 lowering reads only the value, so two thirds of that was unused in
the hashing-dominated verifier circuits.

Make the nibbles optional. Selecting between two constrained bytes yields
a value-only byte, and the paths that need nibbles (XOR, word packing,
comparisons, nibble sums) decompose and range-check them when first asked,
with the same hint and lookup constraints as before. Constructors that
already have nibbles keep them.

On a recursive FRI verifier circuit at 100 queries this takes the
translated circuit from 554k to 390k gates per query.

Share fallible lookup accumulation between proving and verification while preserving public signatures. Reject zero denominators explicitly, including zero-multiplicity messages.

Pack constrained bytes directly in radix 256. Move the auxiliary CUDA host-LDE benchmark into benches and remove experimental timing runners while retaining correctness tests and reference implementations.

Validation: eight focused CPU tests passed with parallel enabled; cargo fmt --all --check, Cargo target metadata, and git diff --check passed.