feat: Enforce max_inflight_requests as a shared limit across ensemble requests

[pskiran1]

This PR changes ensemble scheduling to enforce max_inflight_requests as a shared limit at each ensemble step across all concurrent ensemble requests. This helps control memory usage when upstream steps produce responses faster than downstream models can consume them.

Changes:

• Move per-step in-flight limiters to be owned and shared across all EnsembleContexts at the ensemble model level (EnsembleInfo).
• Update the limiter API from “wait + increment/decrement” to an acquire/release slot model, and wire it into scheduling and completion paths.
• Updates step dispatch to Acquire() a limiter slot before InferAsync() and Release() it on final response completion (or on scheduling failure).

CI and Doc: triton-inference-server/server#8707
Doc: triton-inference-server/common#152

[Copilot]

Pull request overview

This PR changes ensemble scheduling so max_inflight_requests is enforced as a global per-step cap across all concurrent ensemble requests (instead of being enforced per ensemble request), helping bound memory growth when upstream steps outpace downstream consumption.

Changes:

• Move per-step in-flight limiters to be owned globally by the ensemble model (EnsembleInfo) and shared across all EnsembleContexts.
• Update the limiter API from “wait + increment/decrement” to an acquire/release slot model, and wire it into scheduling and completion paths.
• Update configuration/comment semantics to reflect the new global behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/ensemble_scheduler/ensemble_scheduler.h	Updates docs for max_inflight_requests semantics and adds storage for per-step global limiters.
src/ensemble_scheduler/ensemble_scheduler.cc	Implements global limiter allocation and integrates acquire/release into step scheduling and completion.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

Pull request overview

This PR changes ensemble scheduling so max_inflight_requests is enforced as a global per-step limit shared across all concurrent ensemble requests for a given ensemble model, instead of being enforced per-ensemble-request.

Changes:

• Introduces a per-step global limiter (StepInflightRequestLimiter) stored on EnsembleInfo, allocated once per step when max_inflight_requests > 0.
• Updates step dispatch to Acquire() a limiter slot before InferAsync() and Release() it on final response completion (or on scheduling failure).
• Removes the per-EnsembleContext limiter instances and adjusts counter decrement logic to avoid underflow when a step is not actually scheduled.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/ensemble_scheduler/ensemble_scheduler.h	Adds the limiter type and stores one limiter per ensemble step on EnsembleInfo to make the limit global across ensemble requests.
src/ensemble_scheduler/ensemble_scheduler.cc	Moves limiter implementation, wires Acquire/Release into scheduling and completion paths, and removes per-context limiter initialization/usage.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.