Split traces larger than -b instead of dropping them

[adsr]

Supersedes #167, which in turn superseded #157.

Currently phpspy drops traces that are larger than -b, which defaults to PIPE_BUF (4096 bytes). The idea is to avoid interleaved traces in -P mode. The problem is that traces are often larger than 4k in production, so we often bump -b and eat the cost of potentially corrupt traces.

The fix is to split large traces into smaller chunks. As long as each thread outputs each chunk under PIPE_BUF bytes, then readers like stackcollapse-phpspy.pl can reassemble the chunks back into the full trace.

Differences compared to #167:

• Writes happen in a single pass. 167 did one pass to count chunks and another pass to write each chunk as k/n. Here we include a sentinel with each chunk that indicates "incomplete" (~) or "complete" (!).

• There's no fixed reserve for trace_id records. It's calculated dynamically.

• Traces under -b are unchanged -- no trace_id metadata.

• No partial records. If a single frame record exceeds -b, drop it and emit a # truncated = 1 metadata record. We could make it a flag to emit partial records.

• -1 mode is not a special case. 167 disabled chunking in -1 mode.

• The -o fd is open-d then dup-d by other threads under a mutex. This avoids interleaving/corruption in -P -o ... mode. Adds a comment TODO the same for -O and -E. Could be a separate commit.

• Delimiter stripping uses memchr -- more efficient than the existing byte loop. Could be a separate commit.