potentially remove dropna in site find_valid_t0s

Missed this on merge, but re: comments here:

    # 2. Now lets loop over each location in system id and find the valid periods
    # Should we have a different option if there are not nans
    sites = datasets_dict["site"]
    site_ids = sites.site_id.values
    site_config = config.input_data.site
    valid_t0_and_site_ids = []
    for site_id in site_ids:
        site = sites.sel(site_id=site_id)


        # drop any nan values
        # not sure this is right?
        site = site.dropna(dim='time_utc')

I don't think we should be doing dropna here. This will make a block of dates with even 1 missing value discontinuous which I think is less beneficial than using it and filling in the missing timestamp later on (e g one missing point can cost you something like 45 potential t0s with 3h history and 8h forecast). I've used data with about 3% missing, sometimes in considerable chunks, and the model seemed to do fine and not get distracted by nan infills.

There is a greater discussion to be had around how much missing data we allow to be infilled and at what times, but I think this should be done in preprocessing anyway and not here; I'd remove it for now.

The relevant entry point is ocf_data_sampler/torch_datasets/site.py, around the site loop and site.dropna(dim='time_utc'); read how valid_t0_and_site_ids is built after this point. Confirm the change preserves timestamps across missing values and avoids discarding potential t0s, then check the resulting sampling behavior against the issue's 3-hour history and 8-hour forecast example.