fix: get_step returns wrong step + pagination ceiling off-by-one

[vystartasv]

Bugs Fixed

1. get_step() returns wrong step (critical)

InMemoryTaskDB.get_step(task_id, step_id) filters by s.task_id instead of s.step_id:

# BEFORE (bug): always returns first step of the task
step = next(filter(lambda s: s.task_id == task_id, task.steps), None)

# AFTER (fix): returns the correct step
step = next(filter(lambda s: s.step_id == step_id, task.steps), None)

Since all steps in task.steps share the same task_id, the filter matches EVERY step. next() returns the first one — regardless of which step_id was requested. This means GET /ap/v1/agent/tasks/{task_id}/steps/{step_id} always returns the wrong step.

2. Pagination off-by-one (2 locations)

total_pages uses floor division, undercounting when items don't divide evenly:

# BEFORE: 11 items / page_size 10 → total_pages = 1 (should be 2)
total_pages = len(tasks) // page_size

# AFTER: ceiling division
total_pages = (len(tasks) + page_size - 1) // page_size

Fixed in both list_agent_tasks and list_agent_task_steps.

Impact

• get_step: The API returns incorrect step data for any task with multiple steps. Silent data corruption — no error, just wrong results.
• pagination: Clients seeing total_pages=1 for 11 items at page_size=10 miss the second page entirely.