[Fix] Fix distributed POST actor concurrency split

[kaysonyu]

Summary

This PR fixes the per-actor concurrency calculation in distributed POST mode.

The current implementation creates args.num_gpus_per_node POST actors on each alive Ray node, but
it only divides _client_concurrency by the number of nodes when computing per_actor_conc. That
over-allocates concurrency for each actor.

This change keeps the existing actor creation behavior unchanged and only fixes the concurrency split
to use the total number of created actors.

Changes

• Keep the current actor topology unchanged:
  • create args.num_gpus_per_node POST actors per alive Ray node
• Change per_actor_conc from:
  • dividing by len(nodes)
• To:
  • dividing by len(nodes) * args.num_gpus_per_node
• Clamp the value to at least 1
• Update the docstring to match the current actual behavior

Why

Under the current logic, when there are multiple POST actors per node, each actor gets a much larger
concurrency budget than intended.

Example:

• 8 alive nodes
• 8 actors per node
• _client_concurrency = 1024

Before:

• per_actor_conc = ceil(1024 / 8) = 128+

Expected under current actor topology:

• total actors = 8 * 8 = 64
• per_actor_conc = ceil(1024 / 64) = 16

Scope

This PR is intentionally minimal:

• no topology change
• no new config
• no routing change
• no behavior change outside distributed POST concurrency accounting

Testing

• No new committed test in this PR
• Change is limited to concurrency calculation in slime/utils/http_utils.py