fix: make two quiet paths name the molecules they dropped

[isayev]

Cluster A continued (plan: #130). Both findings are one defect in two places: a code path that loses a molecule and says less about it than another path doing the identical thing, so how much the user is told depends on which door they came through.

M1 — the optimizer's SDF reader logged only the all-failed case

mols = [m for m in mols if m is not None]   # a bad record vanishes here
if not mols:
    logger.warning("No valid molecules in input file. ...")   # only this fired

A single bad record among a thousand left the output file shorter than the input with nothing said about which one. Through opt_geometry that is a short SDF, the path returned, and exit 0 — with RDKit's own C++ parse error as the only trace, and that names a file offset, not a molecule.

SPE.calc_spe and ASE/thermo's iter_thermo_records both log per-record for exactly this situation. This was the one reader that did not.

M2 — the parallel embedding path was quieter than the serial one, twice

	serial	parallel (before)
unparseable SMILES	logs the SMILES it could not parse	return [], silent
species embeds nothing	n_written == 0 warning	nothing

A species that embedded nothing was absent from the output, never reached ranking, and did not even produce "No structure converged". use_parallel_embedding is documented as a performance option; it decided whether a lost molecule was reported.

The empty-result warning is raised in the parent, not the worker, and that placement is the point. A message logged inside a ProcessPoolExecutor worker depends on that child's logging configuration; the parent's does not. It also covers both causes at once — unparseable SMILES and every-conformer-clash-rejected both arrive as an empty result — while the child still logs the parse failure, which adds the reason the parent cannot know.

Verification

• Both mutation-verified: restoring the silent filter fails the optimizer test; removing the parent-side warning fails the embed test.
• A third test asserts a molecule that embeds normally is not warned about, so neither new branch can fire on success.
• The optimizer test uses an all-unparseable SDF deliberately, so it returns before any model is needed — the per-record warning under test happens while reading the file.
• 1284 passed, 9 skipped; ruff clean.

Version decision recorded

The next release is 3.0.0, not 4.0.0. No 3.x ever reached a package channel — PyPI's latest is 2.3.1, conda-forge's is 2.3.0 — so skipping to 4.0.0 would leave users wondering where 3.x went, while 2.3.1 → 3.0.0 is a plain major bump that still signals the breaking changes.

Three collisions follow, written down rather than acted on:

1. The v3.0.0 and v3.5.0 tags correspond to no published artifact and need deleting, with v3.0.0 re-created at the release commit. Deleting pushed tags is outward-facing, so it waits for an explicit go-ahead.
2. The CHANGELOG's [3.0.0] and [3.5.0] sections describe milestones that never shipped — recommend relabelling them as never-published rather than deleting, then retitling the unreleased [4.0.0] section.
3. Found while checking the above: the CHANGELOG and the published record diverge in both directions. It carries a [2.2.10] that was never published, and is missing 2.3.0 and 2.3.1, which were. The last CHANGELOG entry matching a real release is 2.2.9.