Skip IDS reads when no rule matches name and occurrence

[chupakobra6]

Related to #10.

find_matching_rules() loads every populated IDS occurrence before checking whether any selected rule can use it. A rule targeting only core_profiles:3 therefore also reads unrelated data, and an unrelated load error can stop that validation.

Select candidates by IDS name and occurrence before calling DBEntry.get(). Keep version filtering after loading: with autoconvert=False, a stored IDS can have a different DD version from the entry's factory. Additional inputs to cross-IDS rules still load on demand, and wildcard rules retain their existing behavior. Document these loading boundaries.

Validation:

• 18 regression cases cover exact read counts, empty/no-match selections, wildcard and occurrence matching, stored DD versions, cross-IDS dependencies, both load-error policies, and real netCDF validation with coverage on/off. On unchanged develop, 13 of these cases fail; all pass with this change.
• Full suite: 203 passed, 2 existing skips on macOS/Python 3.12 with IMAS-Python 2.3.0, and on Python 3.10 with minimum IMAS-Python 2.2.0.
• Integration run of both follow-ups with Speed up assert bookkeeping: cached node paths and lazy stacks #24 (22eb4cd): 287 passed, 3 expected skips on each of Linux arm64/Python 3.10–3.14.
• Black, Flake8, MyPy, Isort, and Sphinx with warnings treated as errors pass.

The benefit is avoiding unnecessary reads in filtered workloads. Generic wildcard rules can still require every IDS; this does not accelerate reads that are actually needed or claim to resolve the complete performance problem in #10.