[cboulay]

• ezmsg-baseproc>=1.15.0: its hash witness no longer holds the first message's axes, which pinned cross-process shared memory and killed the subscriber's channel when the publisher grew it (Cross-process SHM grow kills the subscriber channel (BufferError) when a received message is still referenced ezmsg#272, Hash witness retains the first message's axes, pinning cross-process SHM (causes ezmsg#272) ezmsg-baseproc#17). 1.15.0 itself requires ezmsg 3.10.0b5.
• ezmsg>=3.10.0b5: SHM hardening (#272), coordinate-axis elision across processes, and fingerprints computed when pickling. Note that SHM-delivered messages are now read-only; a scan found no in-place writes to received data in this repo.
• The ezmsg-sigproc>=3.10.0 floor was already compatible with baseproc's stream_dim rename.

Tested locally against ezmsg 3.10.0b5 + ezmsg-baseproc 1.15.0 + ezmsg-sigproc 3.10.0: everything outside tests/integration passes (363 passed, 24 skipped).

Integration tests hang on my machine either way: test_torch_model_unit_system and test_refit_kalman_system both stall in ez.run identically on ezmsg 3.10.0b4 + baseproc 1.14.0, so it's unrelated to this bump. Worth checking whether CI hangs too.