[ianthomas23]

This PR fixes two problems that occur when using pyodide_http in xeus-python kernels using SharedArrayBuffer. Essentially this will allow xeus-python kernels to work in JupyterLite (and xeus-runner) deployments that are hosted with cross-origin headers such that SharedArrayBuffer is enabled.

Fixes are based on similar code in pyodide with the help of Claude.

1. Issue #316, support the dict_converter kwargs in to_js. The converter is passed down the call stack to where it is needed, and dictionary keys and values are recursively converted to JS. Addresses this pyodide_http line:

https://github.com/koenvo/pyodide-http/blob/b5fc794dfd0f6db18304aa9ec4d967e07123b6fc/pyodide_http/_streaming.py#L137

2. Allow python objects to be iterated in JS. Without this there is an error as a JS symbol is used as a key in a dictionary, which can only accept strings, the error is BindingError: Cannot pass non-string to std::string. The fix is to proxy python iterators so that they behave like JS iterators. Addresses this pyodide_http line:

https://github.com/koenvo/pyodide-http/blob/b5fc794dfd0f6db18304aa9ec4d967e07123b6fc/pyodide_http/_streaming.py#L201

Before this PR, if you create a xeus-python JupyterLite deployment and serve it using cross-origin headers (e.g. via npx static-handler --coi dist/), then the python kernel fails to start with:

The following screenshot shows works using this PR, after rebuilding the pyjs emscripten-forge package using this branch, and then the xeus-python package using the rebuild pyjs-rt:

Note the mamba list output shows the use of the 2 locally-built packages.

Fixes #136.