refactor: tighten five annotations that were narrower than the values

[michael-denyer]

Continues the pyrefly burn-down. 77 → 70 entries. Seven entries across five src
modules, each a declaration that did not describe what the code actually passes or holds.

Site	Errors	Fix
core/progress.progress_iterator(iterable: Iterator)	1	Iterable
io/plink.py total= gets int | ndarray	1	int(bed.sid_count)
lmm/likelihood._frozen(data: list[int]) called with tuples	2	Sequence[int]
lmm/loco.py union annotation inside the branch	1	declare before the branch
lmm/special.chi2_sf_batch .astype on a frompyfunc result	1	np.asarray(..., dtype=)
gwas.GWASResult.timing = field(default_factory=dict)	1	default_factory=GWASTiming

A few worth expanding:

• progress_iterator only ever does for i, item in enumerate(iterable), so Iterable
  is what it actually requires. Iterator rejected the range that io/plink.py passes it.
  Widening a parameter cannot break a caller.
• loco.py annotated snps_global_mask: np.ndarray | None on the assignment inside the
  branch that builds it, which made the declared type a union at the point of the very next
  item assignment. Declared before the branch instead.
• _frozen passes data straight to np.array, which takes any sequence; two of its five
  call sites pass tuples.
• GWASResult.timing is the same fix as PipelineResult.timing in refactor: declare real types in pipeline.py and pipeline_config.py #123.

Verification

Two of these sit on numerically sensitive paths — special.py feeds p-value computation and
loco.py gates SNP selection — so this was checked, not argued.

chi2_sf_batch is bit-identical across 13 edge cases (NaN, negative, zero, 1e-300, 1e-12,
0.5, 1, 2, 10, 100, 700, 1e6, inf), comparing against the previous expression evaluated side
by side in the same process:

chi2_sf_batch bit-identical across 13 edge cases: True

Compared (before vs after)	Result
All 4 LMM modes × both backends, .assoc.txt	byte-identical
LOCO: .assoc.txt, pve/pve_se, per-chromosome kinship sums, stats cache	identical
pytest tests/	2243 passed, 10 skipped, 3 xfailed
ruff check / ruff format --check	clean
pyrefly check --baseline on 3.11.13/numpy 2.4.6 and 3.13.5/numpy 2.5.1	0 errors both

What's left in src/

Seven entries, deliberately not in this PR because they are not one-line annotations:

• cli.py:549, gwas.py:189 — backend: str reaching PipelineConfig's Literal. The CLI
  already constrains it with click.Choice, so the type can simply say so.
• cli.py:339 — -gk/-lmm are validated as mutually exclusive and one-required ~70 lines
  above the dispatch, so lmm is not narrowed at the call.
• lmm/runner_numpy_streaming.py ×4 — RunnerTiming misuse (.clear() and
  dict(...) on a TypedDict) plus a _LazySnpMeta assigned to a list | None parameter,
  currently held together by an assert in production code.

Next PR.