Add Claude skill

[JojiiOfficial]

Adds a skill for claude code to run bfb with all context it needs.
It covers all commands (upload, search, scroll, schema), experiment templates, how to download real data and a list of common mistakes.

Validation

Additionally, this PR adds a new subcommand validate, which I found agents were frequently looking for while testing the skill.
It checks each config parses and passes its own validation, then cross-references the upload against the search/scroll configs: vector names, kinds and dimensions, and filter fields against declared indexes.
Obvious things that Qdrant will always reject (e.g. dim mismatch) will lead to an error. Likely misconfigurations, such as random vectors in search but real vectors in the upload config will throw a warning.

Tests

Two agents were given identical bfb tasks: one with the skill available, one without (control).
Note that these numbers fully depend on the initial prompt. The numbers here represent a selected set of smoke tests to verify the skill helps and how much each metric shifts for the specific test prompt.
These numbers don't represent real usage speedups.

Averages across all 5 runs

Metric	With skill	Control (no skill)	Delta
Tool calls	12.6	30.8	2.4× fewer
Wall time	2m53s	5m33s	1.9× faster
Tokens	58.0k	79.9k	1.4× fewer
Recall@10 (dbpedia)	0.612	0.612	identical

bfb source-code accesses (counted only in 2 of the 5 runs)

Task	With skill	Control
Upload-config authoring	0	2 files + 3 greps
Search benchmark	0	0

Per-run detail

Run	Task	Calls (Skill)	Calls (no skill)	Skill time	Ctrl time
1	upload+search (no server given)	27	45	4m55s	8m45s
2	read-only search	11	39	2m33s	5m09s
3	read-only search	10	31	3m12s	5m45s
4	upload-config authoring (offline)	6	18	1m21s	3m55s
5	read-only search	9	21	2m23s	4m10s

Stale Readme section

This PR also removes a stale section about shell completions from the README.md. We currently have it specified twice and my changes remove the old version.