Remove legacy probe construction path for read_imro

[h-mayorquin]

_read_imro_string was the last caller of _make_npx_probe_from_description, which builds probes from scratch using electrode positions, pitch, and stagger parameters. This PR rewrites _read_imro_string to use the catalogue-based approach (the same one read_spikeglx and read_openephys already use) and deletes _make_npx_probe_from_description.

The motivation is to have a single probe construction path. Keeping two independent implementations of geometry construction means they can diverge, and the catalogue-based path is the canonical one that produces contact_ids consistent across all readers.

All three readers now follow the same abstract pattern:

1. Determine probe part number from the input format
2. build_neuropixels_probe(part_number) to get the full catalogue probe
3. Parse the format-specific electrode selection (IMRO table, XML CHANNELS/SELECTED_ELECTRODES)
4. Build canonical contact IDs via _build_canonical_contact_id
5. full_probe.get_slice(selected_indices) to slice to the active electrodes
6. Annotate with format-specific metadata (IMRO fields, ADC sampling, settings channel keys)

The differences between readers are only in steps 1 and 3 (where the part number and electrode selection come from) and what extra metadata each format provides (saved channel subsets, serial numbers, Channel Map ordering).

Contributes to #405.

See #351 for related discussion.

[chrishalcrow]

I think imro_np1000 and imro_np1110 don't have electrode as a key (

probeinterface/src/probeinterface/resources/neuropixels_probe_features.json

Line 2261 in 208a95b

"imro_np1000_elm_flds": "(channel bank ref_id ap_gain lf_gain ap_hipas_flt)",

)
which is what all bank-related logic below (deleted lines 613-621) refer to. So you need to keep that logic in. Annoyingly, we don't have test data for this.

[chrishalcrow]

The bank-y logic should be added to read_spikeglx too, I think.

[alejoe91]

Related: SpikeInterface/spikeinterface#3913

[h-mayorquin]

For Neuropixels 1.0 I codified the fallback in a new function _resolve_active_contacts_for_np1 that encapsulates the electrode = bank * 384 + channel logic. Both read_spikeglx and read_imro now call _get_active_contact_ids which resolves electrode IDs if missing and builds the canonical contact ID strings. This function dispatches to the appropriate resolver based on the IMRO fields present. The IMRO parsing itself stays in _parse_imro_string, which is now a pure parser that returns a dict of IMRO fields without any electrode resolution (the electrode computation that was previously inline there has been extracted into the resolver functions).

For NP1110 I went and checked how it is done in the C++ implementation (IMROTbl_T1110.cpp in the SpikeGLX repo). I threw an agent at it and had it implement the same logic here, which is now in _resolve_active_contacts_for_np1110. I added enough caveats in the function (a warning, a TODO, a comment explaining why we're mirroring the C++ naming) until we get real test data, but I think it should be OK to move forward.

I also inlined all the logic for building the probe into read_imro directly so it is even more like read_spikeglx, instead of relying on a private method that was doing most of the computation.