test(webapp): cover decentralized browser training

[ale18V]

What & why

This PR adds a decentralized webapp end-to-end test for the default MNIST task. A Chrome browser trains with two Node participants over real WebRTC connections. The test also exposed a DiscoJS race that could label outgoing weights with the next aggregation round and leave peers waiting indefinitely; this PR fixes that race. Browser-to-browser coverage is tracked in the stacked #1239.

Technical plan

A dedicated Cypress job starts one MNIST-only test server process with two Node peers, then runs the ordinary Cypress configuration against the webapp. The peers train concurrently with the browser and close when training finishes; the test checks the browser's participant count, epochs, model-sharing rounds, successful completion, and absence of error toasts. DiscoJS captures the aggregation round once for each payload exchange so concurrent arrivals cannot change outgoing message labels.

Deviations from plan

The test uses the existing 24-image CIFAR fixture rather than MNIST samples. It validates image preprocessing, browser-to-Node WebRTC, and model sharing, rather than classification quality or multiclass MNIST accuracy. The CI run also revealed the aggregation-round race, which is fixed here because it prevented reliable completion. The initial harness and separate Cypress config were removed in favor of a simple test-server process, following the review feedback.

Todo

Browser-to-browser Chrome–Firefox coverage is in the stacked #1239.

[JulienVig]

cf. review for #1230, current implementation doesn't test browser-browser collaboration