perf(build): train the upstream-TLS path in the PGO profile (v2 training shape)

Deferred from #967 (2026-08-13 scenario review, option B).

The v1 PGO training matrix drives the gateway against a plaintext local mock, so the TLS client paths (handshake, record encrypt/decrypt on every upstream byte) stay untrained and are laid out as cold code. Production upstreams are HTTPS, making this the largest known divergence between the training profile and real traffic.

What v2 needs:

• an embedded TLS server in bench/pgo-training/trainer (rustls dependency) plus a committed localhost test certificate (clearly marked test-only)
• one or two TLS shapes in the matrix (chat non-stream + one streaming shape)
• a way to measure the benefit: the current probe harness and bench rigs are plaintext, so the gain is unverifiable today — a TLS mock leg for the probe harness is a prerequisite for gating the change
• holdout re-run per #967 hard gate 3 (training-set change)

Rejected for v1 because the trainer would lose its zero-dependency property and the benefit cannot be measured on current rigs.

Start in bench/pgo-training/trainer and review the current PGO training matrix, then inspect the probe harness and bench rigs to understand their plaintext assumptions. The work is done when the trainer supports the specified TLS shapes with a committed test-only certificate, the probe harness can measure a TLS mock leg, and the #967 hard gate 3 holdout rerun is complete.