osi converter: preserve N dimensions over one column across the round-trip (#222)

[ralfbecher]

Closes #222. Builds on #220's extension hook.

Problem

OBML allows N dimensions over one column (grain variants, role-playing via via); OSI is field-centric (one dimension per field). The converter used to break after the first matching dimension on export and build one dimension per field on import, so a second dimension on the same column was dropped silently - both names lost, zero warnings on either side.

Repro (from the issue):

dimensions:
  Order Day:   {dataObject: Orders, column: date, resultType: date, timeGrain: day}
  Order Month: {dataObject: Orders, column: date, resultType: date, timeGrain: month}

OBML -> OSI -> OBML used to yield a single order_date dimension.

Fix

• Export (obml_to_osi): the first matching dimension stays the primary carried on the field; any extras are serialized into an obml_extra_dimensions list extension (name + resultType + timeGrain + format + description + owner + via), and a fidelity warning is emitted (warn, do not raise - conversion still succeeds).
• Import (osi_to_obml): rebuild each preserved descriptor as its own dimension. The collision-key logic is factored into _insert_dimension and reused for the primary and every extra.

Now the same repro round-trips to all three dimensions with their distinct grains, and the export warns rather than silently dropping.

Robustness

Descriptors are opaque foreign-modifiable data (validate_osi does not inspect extension payloads), so each is shape-guarded - must be a dict with a non-empty string name, string props only. A malformed payload (not a list, non-dict items, non-string names) is skipped, never crashes on an unhashable key. This is the same class of guard #228 added for obml_dimension_name, applied here from the start.

Tests

TestMultipleDimensionsPerColumn: all dimensions survive with distinct grains and the same physical column; export warns (not silent); parametrized malformed-payload cases never crash.

177 package + convert-integration tests pass; ruff clean; no new mypy errors (package's pre-existing typing debt unchanged, outside CI's mypy src/ scope).

Upstream note

Whether OSI should natively represent multiple dimensions over one field (cf. LookML dimension_group, Cube granularities) remains an Ossie spec question, tracked in the issue. This PR is the converter-side warn + preserve, independent of that outcome.

Refs #201, #202, #220.