Source side remote view transport

[zhou-yuhan]

Motivation

Remote TP-sliced loads/updates were still hitting the wrong slow path: when a requested
view_id was not already routable, the destination daemon fell back to canonical transport
and reconstructed the TP slice locally. That preserved correctness, but caused destination-
side read amplification, strided repack, and repeated reconstruction across daemons.

We also found lifecycle gaps in the first source-side upgrade path:

• derived view exports were not managed as bounded ephemeral cache entries;
• repeated multi-version workloads could accumulate exports on the source daemon;
• retirement/drain semantics were ad hoc;
• source-side upgrade waiting inherited an internal timeout instead of the caller deadline.

What this PR does

Suppose daemon B wants to fetch tensor views stored on daemon A.

• Implement source-side remote view transport so daemon B can fetch dense view bytes prepared
  by daemon A, instead of reconstructing from canonical bytes locally.
• Add daemon-owned lifecycle management for derived view exports:
  • cache keying by (artifact_id, view_id, device)
  • pending/ready/draining state
  • TTL-scoped reuse
  • pressure-aware eviction
  • ordered retirement
• Add BeginReplicaFetch / EndReplicaFetch so the source daemon tracks real data-plane use:
  • TTL refresh on actual use only
  • active_fetches protection for in-flight transfers
  • drain waits for real relay fetches to finish
• Keep canonical fallback as the compatibility / resource-exhaustion backstop.
• Rework request-budget semantics so source-side upgrade waiting follows the caller deadline,
  not pinned-allocation timeout.
• Add route-kind / fallback-reason observability for create, reuse, refresh, drain, eviction,
  and fallback paths.

Test results

Unit tests

Added focused tests for the new lifecycle manager:

• cache keying + reuse for identical (artifact_id, view_id, device)
• TTL refresh on data-plane use only
• eviction ordering: expired idle first, then oldest idle non-expired, never active/pending

SGLang integration

Validated with the remote relay benchmark harness on two workers.

Remote load (load_weight_remote, tensorcast relay):

• qwen3-14b / qwen3-32b
• tp=1,2,4
• trial=3
• all cases passed

Examples:

• qwen3-14b tp=4: 26s -> 5s -> 5s
• qwen3-32b tp=4: 23s -> 11s -> 12s

Remote update (update_weight_remote, tensorcast relay):

• qwen3-14b / qwen3-32b
• tp=1,2,4
• trial=3
• all cases passed

Examples:

• qwen3-32b tp=2: 40s / 34s / 38s
• qwen3-32b tp=4: 44s / 43s / 43s

Summary

This PR makes source-side remote view transport production-safe:

• source daemon prepares and serves dense views;
• destination daemon stops doing canonical reconstruction on the routed path;
• derived views are reusable but bounded ephemeral cache entries;
• lifecycle, timeout, and fallback semantics are explicit and observable.