feat: per-trace current_context_size in Langfuse usageDetails

[winjer]

Adds a current_context_size key to each generation observation's usageDetails, matching the number the OpenCode TUI "Context" panel shows: input + output + reasoning + cache.read + cache.write of the trace's last assistant message with output > 0.

Non-obvious constraints the diff doesn't show:

• Existing summed usage (input/output/cache keys) is byte-for-byte unchanged; the metric is additive only.
• For Claude Code traces the formula reduces to the 4-field Anthropic sum (reasoning_tokens absent → 0).
• reasoning_tokens is carried through the normalised message.usage seam internally but deliberately NOT leaked into usageDetails.
• Key is omitted (never emitted as 0) when no qualifying assistant step exists, matching the existing omit-don't-zero convention in emit.rs.
• Pi traces unaffected (no usage block).

Full design/plan history in planflow/archive/2026-08-20-current-context-size/.