tp: speed up loading of perfetto exports

Loading a perfetto export spent most of its time copying column data
out of the archive and doing work at EOF which most sessions never use.

- Alias numeric column buffers directly in the (mmap-backed) trace blob
  instead of copying them: Arrow buffers are aligned, so when the
  values are contiguous Slab::FromShared wraps them with copy-on-write
  semantics in FlexVector. Pages are only faulted in when queried.
- Coalesce adjacent slices of the same TraceBlob in
  TraceBlobViewReader so tables spanning ReadTrace chunk boundaries
  stay contiguous (and therefore zero-copy).
- Compute trace bounds lazily via __intrinsic_trace_start/end instead
  of rebuilding a _trace_bounds table, and use the first row of
  ts-sorted tables rather than scanning them.
- Build the static dataframe indexes on first use (thread-safely via
  std::call_once) rather than sorting them all at EOF.
- Skip DeobfuscationTracker work when there are no mapping packets.

On a 285 MB Android trace (532 MB export, 1M slices) export load goes
from 540 ms to 270 ms (raw proto load: 6.2 s).