feat: add configurable routing profiles to llm-d generator

[amito]

Description

PR #274 introduced minimal llm-d manifest generation with hardcoded default EPP routing plugins. This PR makes the routing configuration selectable via named profiles, end-to-end from UI to generated Helm values.

Three routing profiles are available:

• default - Current behavior (prefix-cache-scorer, decode-filter, max-score-picker, single-profile-handler). Backward-compatible: existing API calls produce identical output.
• session-affinity - Adds session-affinity-scorer (weight 3) for conversational/chat workloads where requests from the same session benefit from being routed to the same replica.
• throughput-optimized - Replaces prefix-cache-scorer with load-aware-scorer (weight 3) for batch/throughput-oriented workloads where even load distribution matters more than cache locality.

Changes

• src/planner/configuration/llmd_generator.py - Added ROUTING_PROFILES dict with TypedDict annotations defining plugin configurations for each profile. generate_all() now accepts an optional routing_profile parameter (default: "default"), validates it, and injects the profile's plugins into the template context.
• src/planner/configuration/templates/llmd/values.yaml.j2 - Replaced hardcoded plugin list with Jinja2 loops over plugins and scheduling_plugins template variables. Fixed weight check to use is not none (handles weight: 0 correctly).
• src/planner/api/routes/configuration.py - Added routing_profile field to DeploymentRequest with Literal type validation. Passed through to llmd_generator.generate_all() when stack=llm-d.
• ui/components/deployment.py - Added a "Routing Profile" selectbox that appears when the llm-d stack is selected. Changing the profile resets generated YAML. The selected profile is passed through to the backend API call.
• tests/unit/test_llmd_generator.py - Added TestRoutingProfiles class (4 tests), 2 API endpoint tests, and extracted _extract_epp_config() helper to reduce duplication.

How Has This Been Tested?

6 new unit tests added, full suite passes:

cd src && uv run pytest ../tests/unit/test_llmd_generator.py -v

Tests cover:

• test_default_routing_profile_unchanged - Verifies backward compatibility: default profile produces identical output to pre-profile behavior (same 4 plugins, same scheduling weights).
• test_session_affinity_profile - Verifies session-affinity-scorer appears in both plugins and scheduling config with weight 3.
• test_throughput_optimized_profile - Verifies load-aware-scorer replaces prefix-cache-scorer with weight 3.
• test_invalid_routing_profile_raises - Verifies ValueError for unknown profile names.
• test_deploy_with_stack_llmd_with_routing_profile - API integration test: POST to /api/v1/deploy with routing_profile: "session-affinity" returns 200.
• test_deploy_with_invalid_routing_profile_returns_422 - API returns 422 for an invalid routing profile (Pydantic Literal validation).

UI manually tested: selecting llm-d stack shows the routing profile dropdown; toggling between profiles regenerates deployment files with the correct EPP plugin configuration.

Merge criteria:

• The commits are squashed in a cohesive manner and have meaningful messages.
• Testing instructions have been added in the PR body (for PRs involving changes that are not immediately obvious).
• The developer has manually tested the changes and verified that the changes work

[jgchn]

This is a duplicate of the ROUTING_PROFILES dict keys in llmd_generator.py#L47. I think we should use the keys from that file as the source of truth.

[amito]

Thanks, addressing this in the follow-up commit.

[jgchn]

This list is a duplicate of ROUTING_PROFILES.keys() from llmd_generator.py. Can we import the dict?

[amito]

Thanks, addressing this in the follow-up commit.