While testing pyscript and thus pyodide, it seems that TSNE is broken.
The reason is that we are specifically using 32 bits internally that do not work with pyodide.
I think that we should make sure that TSNE preserve dtype in 64 bits and 32 bits without silently reducing the numerical precision.

The current workaround is to use method="exact" and externalize the initialization and pass a 64 bits NumPy array. Solely this case preserves the computation in 64 bits.