refactor(codegen): shared FFIType IR + single type parser for C/C++/Rust

[gmelodie]

Summary

The three FFI codegen backends each carried their own ad-hoc parser for the Nim type strings in the metadata registry (FFIFieldMeta.typeName, e.g. "seq[Option[int]]"): C recursed via genericInnerType + startsWith("seq["), C++ via genericInnerType, and Rust via hand-rolled t[4 .. ^2] slicing. Three copies of the same leaf-scalar table, three places to fix any new shape — and they had already drifted: nimTypeToRust was missing the int8/int16/uint8/uint16/uint32/byte/float32 cases that C and C++ handle, so those types silently fell through to capitalizeFirstLetter(t) and emitted invalid Rust.

This introduces one structured type model and one parser that all three backends consume.

• New ffi/codegen/types_ir.nim: a recursive FFIType descriptor (ScalarKind + FFITypeKind), the single parseFFIType parser, and a generic renderNative walker driven by a small per-language NativeTypeMap. genericInnerType moves here as a private parser helper (removed from c_cpp_common.nim).
• cpp.nim / rust.nim: nimTypeToCpp / nimTypeToRust are now renderNative(map, parseFFIType(t)). Public signatures unchanged, so callers are untouched. This fixes the Rust scalar drift by construction (the scalar set now lives in exactly one place).
• c.nim: ensureCType switches on FFIType.kind instead of string prefixes; the leaf scalar table is expressed once as scalarCInfoTable keyed by ScalarKind, and leafSuffix/cToken derive from it. Monomorphisation and codec-body emission (emitSeqType/emitOptType/emitStructType) are unchanged — only their input becomes the shared IR.

Codec-body emission and the high-level ctor/method/event wrappers are deliberately left per-backend (they genuinely diverge: Result<T> vs std::future vs flume, unique_ptr vs RAII vs Drop); a later phase could revisit them behind a backend vtable. ftHandle is parsed for completeness but no production caller passes a handleNames set (handles are pre-converted to pointer upstream), so that arm is currently exercised only by the unit tests.

Verification

• New tests/unit/test_rust_codegen.nim pins the drift fix directly: nimTypeToRust over the full scalar set (including the previously-broken int8/int16/uint8/uint16/uint32/byte/float32), seq[byte] → Vec<u8>, ptr T/pointer, nested generics, Maybe == Option, and struct capitalisation.
• New tests/unit/test_types_ir.nim covers parseFFIType directly: the full scalar set, seq[byte]/seq[uint8] → ftBytes, nested seq[Option[seq[int]]], Maybe[T] == Option[T], handle names → ftHandle, ptr T/pointer → ftPtr, unknown name → ftStruct, plus renderNative round-trips.
• Golden regen + diff: regenerated the committed example bindings; the C and C++ outputs for examples/echo and examples/timer are byte-identical to what is on disk, and the Rust output is identical (the examples don't use any of the previously-broken scalar types, so the drift fix changes no checked-in output).
• All codegen unit tests pass under both --mm:orc and --mm:refc (test_types_ir, test_rust_codegen, test_c_codegen, test_cddl_codegen, test_string_helpers, test_meta).

[Ivansete-status]

LGTM! Great simplification indeed! 🙌