[doc] Document the contract of storage-side callables (data_parser, merge_fn) with a tutorial #195

Background

Three KV APIs run a user callable on the SimpleStorage units:

They are powerful but easy to misuse, and their contracts are documented unevenly, in places inaccurately.

Gaps in the current docs

1. put data_parser: "must modify in-place" is inaccurate. The docstrings of kv_put and kv_batch_put say the parser must modify values in place, but the unit requires it to return a dict and raises TypeError otherwise (_apply_data_parser). Modifying in place is allowed, not required.
2. The get and put data_parser look interchangeable but are not. The get docstring calls itself the get-side counterpart of put's parser, but the two differ in arity ((data, args) vs (data)) and in input structure (per-sample lists vs a batched tensor for regular tensor columns). Passing a put parser to kv_batch_get fails on the unit with takes 1 positional argument but 2 were given.
3. merge_fn does not say that old is the stored object. apply_update passes the stored value itself (L255). If a merge_fn modifies old in place, the guarantee that a failed merge leaves storage unchanged no longer holds: samples merged before the failure have already changed.
4. Runtime constraints are not stated in one place. All three run synchronously on the unit's single worker thread, so a slow one delays every other request to that unit. A call that uses one of them is not retried on a storage-unit timeout, and a timeout has an unknown outcome. Callables travel through cloudpickle, so module-level functions must be importable on the units, and tuples inside parser_args arrive as lists.

Expected work

• Docs. One shared contract for storage-side callables in the README, referenced from the docstrings of kv_put, kv_batch_put, kv_batch_get, kv_update, kv_batch_update and their async variants: where it runs, the input structure per API, no in-place modification of stored values, no side effects, keep it light, no retry, serialization, and how errors reach the caller. Fix gaps 1-3 above.
• Dedicated tutorial. A new tutorial/07_*.py that walks through put data_parser (resolving references into content), get data_parser with parser_args (trimming a tensor field and picking items from a list field in one get), and kv_update with merge_fn (appending to a field). Include one in-place example and show why it corrupts stored data.
• Notebook demo. Extend section 11 of tutorial/basic.ipynb ("Lazy Data Parsing with data_parser") with a get-side data_parser demo, and add a kv_update / merge_fn demo.

Related: #182, #191