Additional cleanup in oni_destroy_ctx

[jsiegle]

I needed to add these additional lines to avoid memory leaks on Windows with the acquisition board plugin. Let me know if this makes sense to merge in main.

[jonnew]

Although this commit generally speaks to architectural issues with the calling code, I think that there are some points in here worth thinking about. This library is not thread safe and I have suspicion that the issue here might be resulting from the calling code using different threads for context management a reading frames.

1. Replacing the actions proposed in this comment with debug assertions about non-zero ref counts at time of oni_destroy_ctx call. This seems obvious for guiding devleopment.

2. Freeing the dev_table does seem like a valid addition.

[jonnew]

@aacuevas I would like your opinion here as well.

[jonnew]

After the discussion with @aacuevas I realized my inital review as wrong. My only suggestion is to add an NB:

    // NB: _ref_dec is only called when a new shared buffer is created. We must
    // therefore decrement the active shared buffers here to balance the initial
    // _ref_inc from their creation.
    if (ctx->shared_rbuf != NULL)
        _ref_dec(&(ctx->shared_rbuf->count));

since this is a bit tricky.