feat: trace payloads without truncation

[milanagm]

Closes LFE-16188

Problem

Every captured payload was cut at PI_LANGFUSE_MAX_CHARS (default 20000), so a long prompt, a long answer or a large tool result reached Langfuse shortened, and the trace no longer showed what actually happened.

What this does

Removes PI_LANGFUSE_MAX_CHARS, truncateText() and TruncationMeta, so these all go out at full length:

Value	Observation field
event.prompt	turn root input.content, and the first generation's input
state.lastAssistantText	turn root output.content
assistant text on message_end	generation output.content
tool arguments	Tool: span input
tool result	Tool: span output
compaction summary	Compaction output.content
branch summary	Branch Summary output.content

Also gone:

• The metadata that only existed to report a cut: user_text_meta, assistant_text_meta (root and generation), args_meta, output_meta, summary_meta (compaction and branch summary).
• A second, harder limit that was not configurable: outText.slice(0, 4000) on pendingToolResults, the tool results that appear as the next generation's input. They were cut twice — first to 20000, then to 4000.
• The node:crypto import, which only served the truncation marker's sha256.

Tests

pnpm test: 60 passing (the two truncateText unit tests went with the function), and tsc --noEmit clean. Verified against all three CI matrix versions — pi 0.83.0, 0.84.3 and 0.85.0.