refactor(koalabear): make Fp an int subclass and tighten field surface

[tcoratger]

Summary

• Switch Fp from a HAS-A wrapper (self.value: int) to an IS-A scalar (class Fp(int, SSZType)) matching the BaseUint primitive pattern. Drops the .value indirection at every call site.
• Replace explicit Fermat exponentiation in inverse() with the stdlib pow(a, -1, P) form (modular inverse via extended Euclidean).
• Move from bare ValueError / TypeError to the project's SSZ exception hierarchy (SSZTypeError, SSZSerializationError, SSZValueError).
• Strict typing on arithmetic with reverse-operator guards so int + Fp raises instead of silently returning a plain int.
• Loose equality (Fp(5) == 5 returns False, doesn't raise) so collection membership checks still work; __hash__ mixes in the type to keep Fp(5) and 5 distinct in sets.
• Add a Pydantic __get_pydantic_core_schema__ hook: instances pass through, raw ints are strict-validated against [0, P), JSON round-trips drop the subtype.

Migration

• xmss/message_hash.py, poseidon1/permutation.py, and the consensus_testing poseidon fixture drop their fp.value reads in favor of int(fp).
• types/collections.py collapses the now-dead IntFieldElement protocol into a plain isinstance(item, int) check that also covers BaseUint.

Test coverage

• tests/lean_spec/subspecs/koalabear/test_field.py: 11 -> 55 tests, branch coverage 84% -> 100%.
• New parametrized groups cover __new__ rejection of non-ints, modular reduction at boundary inputs, reverse-operator rejection, __pow__ edges (zero base, modular exponent), inverse multiplicative identity, full equality matrix, hash-mixing, Pydantic schema lift/reject/round-trip, and the P - 1 boundary in serialize/deserialize.

Test plan

• uv run pytest — full suite (2956 tests pass)
• uv run pytest tests/lean_spec/subspecs/koalabear/ --cov=src/lean_spec/subspecs/koalabear/field --cov-report=term-missing — 100% coverage
• `just check` — ruff lint, ruff format, ty typecheck, codespell, mdformat all pass

🤖 Generated with Claude Code