docs: every install instruction we ship now names [all], not a partial extra

Operator directive after the outage: "ソースコードにそこがオールじゃないと壊れるので、
オールに直しといてくださいね".

25 install instructions across 17 files told the reader to install a PARTIAL
extra -- scitex-cards[mcp], scitex-todo[mcp], [web], [postgres]. All now say
[all].

THE ONE THAT MATTERS MOST is src/scitex_cards/_skills/scitex-cards/
21_fleet-mcp-rollout.md, which is the FLEET-FACING rollout instruction and said:

    uv pip install -U 'scitex-cards[mcp]>=0.7.1'

That is the exact pin the agent-container defs carry, and it is very plausibly
where they got it. If so, the outage did not originate in sac's recipe -- it
originated in my documentation, and sac followed it correctly. I cannot prove
the causal link and am not claiming it as fact, but a package whose own rollout
guide prescribes the failing pin does not get to call the failure someone
else's.

Error-path hints are included, not just prose. When _backend_connect cannot
find psycopg it now says `pip install 'scitex-cards[all]'` rather than
[postgres]: a remedy is an instruction, and a remedy that under-installs is
worse than no remedy because the reader believes they have complied.

TWO PLACES THE MECHANICAL REWRITE WENT WRONG, both caught and repaired:

  - test__help_announces_scitex_cards.py carried a comment explaining what the
    hint USED to say. Rewriting the quoted-historical string turned it into
    "the hint used to say <the current value>", which is nonsense. Restored,
    and extended to record why the partial half mattered.
  - Three skip reasons became "`scitex-cards[all]` extra absent", which is
    simply untrue: fastmcp comes from [mcp]; [all] is the REMEDY, not the
    missing extra. Reworded to state the fact and the remedy separately.

That is the hazard of keying a rewrite on a bare string -- prose about a value
matches the same pattern as the value.

Verified: 25 partial-extra install instructions before, 0 after, and the tests
that assert on hint text pass against THIS worktree. The first run of them
failed, and the reason is worth recording: it imported the INSTALLED
scitex_cards rather than the worktree, so it was asserting my new expectation
against old code. Re-run with PYTHONPATH pinned and __file__ printed: 12 passed.
