Optimize keccak() for raw bytes inputs

[sanjaysudagani]

Summary

This adds a direct fast path in eth_utils.keccak() for raw bytes, bytearray, and memoryview inputs when no hexstr or text argument is provided.

Today those inputs always go through to_bytes(), which means the common raw-bytes case still pays the conversion-argument validation and dispatch cost before hashing.

What changed

• add a raw-bytes fast path in eth_utils.crypto.keccak
• preserve the existing conversion path for hexstr, text, ints, bools, and conflicting argument validation
• add dedicated keccak() tests, including a regression test that verifies raw bytes input no longer calls to_bytes()

Validation

• pytest tests/core/crypto-utils/test_crypto.py tests/core/conversion-utils/test_conversions.py -q
• pre-commit run --files eth_utils/crypto.py tests/core/crypto-utils/test_crypto.py

Benchmark

On this machine, the common keccak(bytes) path went from about 2.90us per call to about 2.33us per call for a 32-byte payload, reducing overhead over the raw backend from about 29% to about 4%.

Closes #95