[py] Adopt selected ruff rule families surfaced by ruff 0.16

## Background

Ruff 0.16 expanded its **default** rule set from 59 rules to 413. Because `py/pyproject.toml`
only set `extend-select`, the project silently inherited that new default and
`bazel run //py:ruff-check` reported **192 errors** on an unchanged tree.

PR #17937 resolved the immediate breakage by changing `extend-select` to `select`, so the
enforced set is now stated explicitly rather than inherited. That was deliberately a
no-op on enforcement: the resolved rule set is identical to what ruff 0.15.21 enforced
(206 rules), plus 3 new `RUF` rules the project already opts into and passes.

This issue tracks adopting the parts of the new default that are actually worth having,
deliberately and in tranches, rather than all at once via a dependency bump.

## What the 192 findings look like

| Family | Count | Notes |
|---|---|---|
| `B` (bugbear) | 35 | High value, but 17 of 19 `B018` hits are a false positive (see below) |
| `BLE` (blind-except) | 30 | Almost all deliberate best-effort cleanup |
| `SIM` (simplify) | 28 | Mostly cosmetic; `SIM115` is worthwhile |
| `PLW` (pylint warn) | 16 | Contains 3 real bugs |
| `S` (bandit) | 15 | 14 of 15 are `S110` try/except/pass |
| `TRY` | 11 | Opinionated exception style |
| `FURB` | 10 | Modernization, mostly auto-fixable |
| `PYI` | 9 | `PYI041` redundant-numeric-union is real |
| `PIE` | 8 | Contains a real bug (`PIE796`) |
| `EXE` | 6 | Shebang / executable-bit mismatches |
| `C4` (comprehensions) | 6 | Mechanical and safe |
| `G` (logging-format) | 5 | `G201` is reasonable |
| `PLR` | 4 | All `assert 1 == 0` "should have thrown" idioms |
| `DTZ` | 4 | All naive `datetime.now()` used for elapsed-time deltas |
| `PERF` | 2 | Trivial and real |
| `W`, `PLE`, `FLY` | 3 | One each, all worth fixing |

## Recommendation

### Tranche 1 — adopt now (~37 findings, mostly auto-fixable)

`W`, `PLE`, `FLY`, `PERF`, `C4`, `EXE`, `PIE`, `FURB`, `PLR`

Low noise, high signal, and largely `--fix`-able. `PLE` and `PIE` each catch a genuine
bug (see the companion bugs issue). `PLR`'s four hits are all `assert 1 == 0,
"should have thrown an exception"`, which should become `pytest.raises` regardless.

### Tranche 2 — adopt with scoping (~44 findings)

`B`, `PYI`

`B` is the highest-value family here (`B006`, `B008`, `B015`, `B017` all flag real
problems), but it needs one carve-out: **17 of the 19 `B018` "useless expression"
hits are `alert.text` inside a `with pytest.raises(...)` block** — the idiomatic way to
assert that a property access raises. Adopting `B` without ignoring `B018` under
`py/test/**` would mean sprinkling `_ = alert.text`, which is strictly worse. Suggest:

```toml
[tool.ruff.lint.per-file-ignores]
"py/test/**" = ["B018"]
```

### Tranche 3 — discuss (~44 findings)

`SIM`, `TRY`, `G`

Defensible either way. `SIM115` (context manager for file opens) is the one with real
substance; the rest is style. Worth a maintainer opinion before spending the churn.

### Decline for now (~45 findings)

`BLE` (30), `S` (15), `DTZ` (4)

- `BLE001` and `S110` are the same story: broad `except` / `try-except-pass` in browser
teardown and test cleanup, where swallowing is usually deliberate. Adopting means 45
individual judgment calls plus permanent `# noqa` churn, for little signal.
- All four `DTZ005` hits are `datetime.now()` used to measure elapsed time in
`w3c_interaction_tests.py`. Naive datetimes are correct for deltas; the rule is
aimed at a problem this code doesn't have.

## Note on mechanics

Adopting a family means adding it to `select` in `py/pyproject.toml`. Keeping `select`
(rather than reverting to `extend-select`) is what prevents the next ruff release from
re-introducing this whole situation.

## Cross-binding impact

None. This is Python lint configuration only.