Record per-call latency in ai_metering_usage to surface provider response times

Problem/Motivation

ai_metering_usage records tokens and cost for each AI call but not
how long the call took. Latency is a key signal for comparing providers and
models: a cheaper model that is consistently slow may not suit an editorial
workflow, and a latency spike can indicate a provider issue before it shows up
in error logs.

The AI Initiative issue https://git.drupalcode.org/project/ai_initiative/-/work_items/3586519 lists latency as an explicit requirement alongside
tokens and cost.

Steps to reproduce

1. Make AI calls through any configured provider.
2. Open the AI Usage Log at /admin/reports/ai-metering/log.
3. There is no response time column; latency data is not recorded anywhere.

Environment

• Drupal version: 11.3.11
• ai_metering version: 1.0.0-alpha1
• AI provider module + version: any (e.g. ai_provider_anthropic 1.2.2)
• AI feature module tested: N/A

Proposed resolution

Record a start timestamp in AiPreGenerateSubscriber (the event that
fires immediately before the provider call) and make it available to the
post-generate path. In AiPostGenerateSubscriber, compute the delta
in milliseconds and write it to a new nullable latency_ms integer
column on ai_metering_usage.

Deliver the column via hook_update_N(). Default to NULL for rows
recorded before the update, so the dashboard can distinguish "not yet measured"
from "measured at zero". Surface latency in the AI Usage Log view as a new column
and in the dashboard breakdown as average latency per provider/model.

Remaining tasks

• Add nullable latency_ms integer column to ai_metering_usage via hook_update_N() and hook_schema().
• Capture start time in AiPreGenerateSubscriber and pass it to the post-generate path.
• Write the delta (ms) to latency_ms in AiPostGenerateSubscriber.
• Add latency_ms to the AI Usage Log view.
• Surface average latency per provider/model in the dashboard breakdown.
• Add a kernel test asserting latency_ms is populated and greater than zero.
• Review.

User interface changes

Yes. The AI Usage Log gains a latency column; the dashboard breakdown shows average response time per provider and model.

API changes

N/A

Data model changes

Yes. A nullable latency_ms integer column is added to ai_metering_usage via hook_update_N(). Existing rows receive NULL.

AI assistance

N/A