feat: add CLI command for PSM table assembly from Winnow output

## Context

The [INFlow pipeline](https://github.com/instadeepai/INFlow) (nf-core/denovoproteomics) has an `ASSEMBLE_PSM_TABLE` step that merges per-sample Winnow outputs into a single experiment-level PSM table. This is currently implemented as an inline Python script in the Nextflow module ([`modules/local/assemble_psm_table/main.nf`](https://github.com/instadeepai/INFlow/blob/dev/modules/local/assemble_psm_table/main.nf)).

Since the inputs are Winnow's own outputs, this should be a Winnow CLI command (e.g. `winnow assemble`) rather than pipeline-embedded logic.

## What the inline script does

1. Reads all per-sample `*_preds_and_fdr_metrics.csv` files (Winnow FDR output)
2. Reads all per-sample `*_metadata.csv` files (Winnow metadata output)
3. Merges each pair on `spectrum_id` (inner join)
4. Extracts `experiment_file` from `spectrum_id` (format `filename:scan`) or uses `experiment_name` column
5. Concatenates all samples into one experiment-level table
6. Drops Winnow internal feature columns that contain serialised arrays (e.g. `mz_array`, `intensity_array`, `token_log_probs`, `theoretical_mz`, etc.) — these break downstream TSV consumers
7. Writes `psms.tsv` (tab-separated)

## Proposed CLI

```bash
winnow assemble \
--fdr-results sample1_preds_and_fdr_metrics.csv sample2_preds_and_fdr_metrics.csv \
--metadata sample1_metadata.csv sample2_metadata.csv \
--output psms.tsv
```

Or with directory input:
```bash
winnow assemble \
--input-dir results/ \
--output psms.tsv
```

## Acceptance criteria

- [ ] CLI command that takes Winnow FDR + metadata CSVs and produces a merged PSM table
- [ ] Drops serialised array columns by default (with `--keep-arrays` flag to override)
- [ ] Tab-separated output by default
- [ ] Handles missing `experiment_name` column gracefully (falls back to `spectrum_id` prefix)
- [ ] Works with both single-sample and multi-sample inputs

Start at the Winnow CLI entry point and use modules/local/assemble_psm_table/main.nf as the behavioral reference for the existing inline Python script. Run the proposed winnow assemble command with CSV inputs and verify merged tab-separated output, array-column removal, the --keep-arrays override, directory input, and single- versus multi-sample handling.

Written by the indexing model from the issue text.