perf(flags): Avoid BigInt allocations in identity hashing

[jaydgoss]

Summary

Cut the digest-to-remainder step from about 61 ns to 2 ns by replacing BigInt with fixed-width arithmetic. These are isolated local release-mode measurements on an Apple M4 Max, not full feature-check or request timings.

This removes a repeated cost in #184: new flags calculate a feature-specific hash on each check instead of reusing the cached identity-only hash.

Replace the BigInt conversion in FeatureContext::compute_id() with fixed-width arithmetic. This change preserves the SHA-1 input and the resulting bucket IDs.

• Read the 20-byte digest as five big-endian 32-bit words.
• Use a plain loop: multiply the running remainder by 2³², add the next word, then take the remainder modulo 1_000_000_000. Reducing after each word gives the same result as converting the whole digest first.
• Each intermediate value is below 1_000_000_000 × 2³² < 2⁶², so it fits in a u64 without heap allocation.
• Move num from runtime dependencies to test dependencies.
• Compare 10,000 generated identities of different lengths with the original BigInt calculation through the existing FeatureContext::id() API.

This draft is a possible future optimization for #184, not its base or a release dependency. #184 keeps the existing BigInt calculation and only removes the temporary feature-name prefix allocation. This PR does not add feature-specific hashing, cutoff handling, or a new cache. The existing ID cache is unchanged, so repeated checks that already use a cached ID do not gain this optimization.

Validation

• cargo test -p sentry-options --lib: 97 tests passed.
• cargo fmt --all -- --check: passed.
• cargo clippy --workspace --all-features --tests -- -D clippy::all: passed.
• git diff --check: passed.