feat(typespec): parameterized backend_type fidelity + raise on invalid (item 54, gap 1)

[discreteds]

Summary

Item 54, PR-1 (gap 1 + validation strictness): the canonical resolver _resolve_field_native silently dropped parameterized backend_type strings on Polars/Narwhals/PyArrow (Ibis/Pandas were already correct), and silently absorbed invalid backend_type strings into a canonical fallback.

This PR:

• Adds core/dtypes/_paramstring.py — a restricted AST-validated parser for constructor-call-style dtype reprs (Polars/Narwhals str(dtype) output), shared by both targets. Closed whitelist namespace; adversarial-tested (no eval of unvalidated input).
• Upgrades parse_type_string on Polars and Narwhals (safe-eval parser) and PyArrow (bounded bracket/paren regex family: timestamp/duration/time/decimal) — parameterized dtypes like Datetime(time_unit='us', time_zone='UTC'), Decimal(precision=38, scale=10), List(Int64), Array(Int64, shape=(5,)) now round-trip to real native dtypes instead of canonical fallback.
• Adds InvalidBackendTypeError and wires it into _resolve_field_native: a non-empty, non-None backend_type the target cannot parse now raises (names field + string + target) instead of silently degrading. None/"" fall through unchanged (item 53's ANY→STRING path preserved).
• Flips TestInlineReadItem54Deferred → TestInlineReadParameterizedBackendType (the deliberate flip item 53 left for item 54).

Spec: 2026-08-16-converters-resolver-dtype-fidelity-design.md (rev 2, GLM-5.2 adversarial review folded). Plan tasks 1–6, RED-first TDD, one commit per task.

Verification

• Full suite: 22935 passed, 0 failed
• hatch run ruff:check: clean
• hatch run mypy:check: 0 new errors (98 pre-existing, all import-untyped/fixture-class errors present at baseline; touched files contribute only pre-existing pyarrow import-untyped)
• Grep-confirmed: no production FieldSpec(backend_type=...) string literal exists; only test usages (all parseable or intentional raise cases)