Register and ship generic-format adapters (`features-json`, `features-csv`, `features-tsv`, `bed`)

[dlrice]

Context
features-json, features-csv, features-tsv, and bed are declared as canonical names in the KnownAdapterName union in src/schema/types.ts, and the file-extension shorthand in src/schema/normalize.ts infers them (.csv → features-csv, .tsv → features-tsv, .json → features-json, .bed → bed). But no runtime registration of these adapters exists — registerBuiltinAdapters() appears only in doc comments in src/schema/registry.ts. As a result, a config that relies on the inference (e.g. data: "./x.csv") infers adapter: features-csv and then fails at validate/load time with "Unknown adapter" because no such adapter is registered.

If we want the bring-your-own-CSV authoring workflow to actually work end-to-end, these adapters need a real implementation.

Task
Implement and register the four generic-format adapters so they are available out of the box on the <protvista-uniprot> element.

Scope:

• Add src/schema/adapters/features-json.ts — parses an array of feature-shaped records already in the expected shape. Happy path + malformed input (non-array, wrong field types) unit tests.
• Add src/schema/adapters/features-csv.ts — parses CSV with header row type,start,end,description[,score]. Uses the existing d3-dsv dependency (no new dep).
• Add src/schema/adapters/features-tsv.ts — same column convention, tab-separated. Shares the parser core with features-csv.
• Add src/schema/adapters/bed.ts — standard BED (tab-separated). At minimum BED3 (chrom/start/end); ideally up to BED6 for score + strand.
• Implement registerBuiltinAdapters(registry) in src/schema/registry.ts and call it from the default registry factory so the adapters are pre-registered before any config loads.
• Reinstate the deleted spec material in specs/config-approach.md: Example 4 (bring-your-own CSV + transform pipeline), the Edge Cases row for malformed generic-format-adapter input, and the corresponding acceptance criteria.
• Each adapter has a unit test covering happy path and at least one malformed-input case, producing a descriptive error that identifies the offending row/column.

Notes:
The payload shape these adapters produce should match what the resolved semantic-kind adapter consumer expects (feature-shaped records with type, start/begin, end, description, optional score). Validate against the existing nightingale-track-canvas component to confirm tracks render without any per-track glue code from the author.