feat(esm_catalog): PR-A1f — paleo extension + experiment classification

[siligam]

Part of the esm_catalog decomposition, per the PR-A1a foundation plan. Follows #1504 (A1b core), branches from release. Fields 21–27.

Adds src/esm_catalog/paleo.py:

• add_paleo_extension(item, paleo_config, paleo_year, reference_year) — sets paleo:year, paleo:display, paleo:reference_year, paleo:epoch, paleo:period. Year priority: explicit param → paleo_config["reference_year"] → extreme year (< 0 or > 9999) parsed from the item's own datetime strings. No-op for non-paleo items.
• add_experiment_type(item) — classifies every item as paleo/control/historical from the paleo:year property or the item's own datetimes.

Plan-mandated fixes vs the old pr-a1a version:

• _add_experiment_type no longer reads nml:echam:runctl:dt_start (fragile assumption that echam is always a component, plus an ordering dependency on the namelist extension) — there's a regression test asserting the namelist property is ignored.
• paleo:years_bp is only added when experiment_type == "paleo", not for every dated item.
• Dead if TYPE_CHECKING: pass removed; uncalled detect_paleo_simulation removed.

CollectionContext gets an optional paleo_config field (the general.paleo config section as a plain dict); make_item wires both functions in.

Also adds configs/stac-extensions/paleo/v1.0.0/schema.json (custom extension) and tests/test_esm_catalog/test_stac_paleo.py.

Note: like #1511/#1512, this touches the same two spots in item.py — whichever of the three merges last will need trivial conflict resolution (keep all import/call lines).

Test plan

• pytest tests/test_esm_catalog — 65 passed
• New tests cover: no-op behavior, explicit/config/deep-time-string year resolution, epoch/period, Ma/ka/BCE/CE formatting, ISO year parsing edge cases, all three experiment_type classes, years_bp only for paleo, namelist-property independence, make_item wiring, and jsonschema validation (schema also probed to reject bad enum values and non-integer years)

[siligam]

Same silent-null-write class as the roles thing on #1506, just spelled with if key in dict: + direct index instead of .get(key, default) — so a .get(..., [ grep won't catch it. A bare epoch: (or period:) under general.paleo in YAML is present-but-None, so "epoch" in paleo_config is True and "paleo:epoch": null gets written into the item. If the paleo schema types these as strings it's schema-invalid, and nothing in src/esm_catalog calls .validate(), so it ships unchecked (see #1506 discussion + repro: https://gist.github.com/siligam/58cad44b97c831c2f41d22d681c3ddbd).

Cheap fix: gate on truthiness instead of presence — if paleo_config.get("epoch"): / if paleo_config.get("period"):.

[siligam]

Fixed in 19a5faf — gated on paleo_config.get("epoch")/.get("period") truthiness, plus a test_bare_epoch_period_do_not_write_null regression test. Full paleo suite green (26 passed).