[adeebshihadeh]

Alternative to #716: same goal (drop Cython), using nanobind instead of the raw CPython C API. Same public API, module names, and small behavior changes as #716, so the two can be compared directly.

nanobind's stable ABI needs Python >= 3.12, so wheels are one cp311 wheel plus one cp312-abi3 wheel per platform. The abi3 wheel passes the full test suite on 3.12, 3.13 and 3.14, and the cp311 wheel on 3.11.

Benchmark (in-process, ns/op, best of 3, pinned core)

	Cython (master)	C API (#716)	nanobind (this PR)
send + recv, 64 B	2706	2609	2697
send + recv, 1 KiB	2815	2718	2806
send + poll + recv	2957	2731	2821
receive(non_blocking=True), empty	56	21	31
all_readers_updated()	19	17	24

Cross-process 1 KiB ping-pong is identical within noise (~160k msgs/s for all three).

vs #716

• ~300 lines of bindings instead of ~980
• adds nanobind as a build dependency, and builds its runtime into a static lib from SCons
• stripped .sos are ~560 KB total vs ~200 KB for Replace Cython with the CPython limited API (abi3) #716 (~480 KB with Cython)
• two wheels per platform instead of one (3.11 can't use nanobind's abi3)
• Intel macOS wheels need 10.14+ (nanobind uses aligned allocation); Replace Cython with the CPython limited API (abi3) #716 keeps the 10.9 default
• bool arguments are strict (True/False only), where Cython and Replace Cython with the CPython limited API (abi3) #716 accept any truthy value