Add experimental multi-threaded ondemand stream parsing [research] - #2788

Introduce simdjson::experimental::parse_many_parallel, which scales the
stage-2 (value materialization) step of iterate_many across many threads.
A single dispatcher thread carves the input into document-aligned slices and
feeds a pool of worker threads through a bounded queue. Each worker owns its
own ondemand::parser and its own output vector, so no locking happens on the
hot path; extracted values are returned as one vector per worker ('shards').

Two stream formats are supported, each sliced at a delimiter byte that cannot
occur inside a JSON value, so a document is never split and boundaries are
found with a single memchr:
 - whitespace_delimited (ndjson): cut just after a newline
 - json_sequence (RFC 7464): cut just before a record separator (0x1E)
Because the dispatcher stays cheap, throughput scales with the parser threads
(compute bound) or up to memory bandwidth (light workloads).

This targets the case the built-in two-thread iterate_many cannot help with:
when stage 2 dominates, overlapping stage 1 of the next batch only buys a
small constant factor, whereas running many parser threads scales with cores.

Adds benchmark/bench_parallel_stream.cpp comparing serial, built-in
two-thread, and parallel(N) over the amazon_cellphones corpus and a
number-heavy synthetic set, plus a format-coverage check that confirms the
parallel document count and checksum match the serial parse for each format.
Header-only; falls back to single-threaded when SIMDJSON_THREADS_ENABLED is
not defined.