add synthetic data source for running without trodes

[jl33-ai]

right now the only way to run the pipeline is against a live Trodes acquisition. that gate makes a few useful things harder than they need to be:

• smoke testing a fresh install (does pip install actually work end to end?)
• developing or debugging the decoder loop on a laptop
• CI / regression tests
• letting outside contributors poke at the codebase before they have a rig stood up

this PR adds an in-process synthetic data source so the whole MPI pipeline can run without any hardware.

• realtime_decoder/synthetic.py
  • SyntheticDataReceiver is a drop-in for TrodesDataReceiver. emits poisson spikes per ntrode, white-ish LFP, and a triangle-wave position walk along a single linear segment. non-blocking __next__ semantics so the polling main loops in encoder/decoder/ripple work unchanged.
  • SyntheticClient is a drop-in for TrodesClient. same surface used by the supervisor and stim decider. fires startup/termination callbacks on wall-clock timers and records would-be ECU shortcut messages instead of sending them.
• runscript.py picks the data source via config['datasource'] ('trodes' default, 'synthetic' new) using a small factory. no change for existing configs.
• config/demo_synthetic.yml is a 5 rank demo wired to the synthetic source (1 supervisor, 1 ripples, 1 encoder, 1 decoder, 1 gui). output to /tmp, no hardware paths, auto-terminates after 30s.
• README.md adds a 'running without acquisition hardware' section so the install-and-try path is discoverable.

caveat: not biologically realistic. spikes are poisson, marks gaussian, the synthetic animal walks back and forth on one segment. the point is to exercise the plumbing, not to validate decoding accuracy. the stim decider's ECU shortcut messages become no-ops in synthetic mode, which is intentional for a demo.

to try it:

mpiexec -np 5 python -u runscript.py config/demo_synthetic.yml