Accumulate the conversation across multi-step tool-calling steps

[cwurm]

create_next_step_options was called with initial_options as the base, so every step's request contained only the original prompt plus the latest step's tool calls and results — the tool exchanges of all earlier steps were dropped.

The model could therefore never see more than one tool result at a time: asked for something that requires several tool calls, it kept re-requesting the results it had just lost until max_steps was exhausted, without producing an answer.

Observed in clickhouse-client AI SQL generation, where a question needing two table schemas made the model alternate between the same two get_schema_for_table calls until the step limit:

:) ?? query all nodes for their partitions with most active parts (only show one row per node, ...)
•  Starting AI SQL generation with schema discovery...
─────────────────────────────────────────────────
🔧 Calling: list_databases [toolu_01...]
✓  list_databases completed
🔧 Calling: list_tables_in_database [toolu_01...]
   └─ Args: {"database":"system"}
✓  list_tables_in_database completed
🔧 Calling: get_schema_for_table [toolu_01...]
   └─ Args: {"database":"system","table":"parts"}
✓  get_schema_for_table completed
🔧 Calling: get_schema_for_table [toolu_01...]
   └─ Args: {"database":"system","table":"clusters"}
✓  get_schema_for_table completed
🔧 Calling: get_schema_for_table [toolu_01...]
   └─ Args: {"database":"system","table":"parts"}
✓  get_schema_for_table completed
🔧 Calling: get_schema_for_table [toolu_01...]
   └─ Args: {"database":"system","table":"clusters"}
✓  get_schema_for_table completed
🔧 Calling: get_schema_for_table [toolu_01...]
   └─ Args: {"database":"system","table":"parts"}
✓  get_schema_for_table completed
🔧 Calling: get_schema_for_table [toolu_01...]
   └─ Args: {"database":"system","table":"clusters"}
✓  get_schema_for_table completed
─────────────────────────────────────────────────
•  ⚠️  No SQL query was generated: the schema exploration step limit was reached (ai.max_steps = 10)

Base the next step on current_options, which carries the accumulated conversation, so each step sees all previous tool calls and results.

Added a regression test that drives MultiStepCoordinator::execute_multi_step with a scripted generate function (no network) and asserts that the third step's request still contains the first step's tool exchange.

[cwurm]

Closing: this came out of working through pushing some older local changes upstream - the submodule pointer in ClickHouse/ClickHouse is still at ab06ef7, from before this was fixed on main in #40, so the bug was reproduced and fixed independently against the stale base. The remaining useful piece is bumping the submodule pointer in ClickHouse/ClickHouse to current main, which will happen separately.

[cwurm]

ClickHouse/ClickHouse#113959