test(sdk): run the Python client against a real server

[glatinone]

test(sdk): run the Python client against a real server

Every test in this package mocked the transport, and that is how forget on the
async client shipped without the archiving PATCH the protocol requires: the mock
answered the DELETE and never modelled archived -> deleted, so a method that
could not work against a real server passed its tests. A mock answers whatever the
test tells it to. Only a real server disagrees.

sdk/python/tests/test_live_client.py is ten cases against a running server:

• the round trips, including that reading a cell bumps its access count
• forget on both clients, which is the regression this file exists for
• the access rule as a client sees it: an agent the memory was not shared with gets
  an empty result, not an error
• the protocol error codes as the SDK parses them, including 403 ACCESS_DENIED
  for an unknown id (never 404, per spec §8.4) and 422 VALIDATION_ERROR for a
  rejected body - which also pins the previous fix from the client side
• paging with offset

Proven to fail, not just to pass. With the old broken forget planted back,
the live test fails with the server's own words: INVALID_TRANSITION: Cannot delete cell with status 'active'. Archive it first. Restored, it passes. A test that has
never failed is a test nobody has checked.

CI runs it, and does not let it skip itself:

• The sdk job installs and starts the reference server and runs the live file with
  AMP_REQUIRE_LIVE_SERVER=1, which turns "no server reachable" into a failure -
  the same guard the storage job uses for PostgreSQL.
• Not on the 3.10 leg, and by an explicit if rather than a skip: the reference
  server's own floor is 3.11, so that leg has nothing to talk to. A skip would read
  as a pass.
• Locally the file still skips with a message naming AMP_TEST_URL, so the package
  tests on a laptop with nothing running (verified: 10 skipped without a server,
  10 failures with the flag set and no server).

[glatinone]

Landed on master in the v0.1.0 chain: the branch was fast-forward merged as part of b940905..92b88ee and released as v0.1.0. Closing so the open list matches reality - the commits are in master, and the tag points at them.