fix: skip null and NaN line x instead of failing, and stop zooms and selections past an integer dtype's range from returning nothing - #152

An ungrouped line on an in-memory frame with a null or NaN in x returned 500 for every request. The line now skips those rows, as histograms do.

• check_line_x accepts nulls at the start and NaN at the end of a sorted x (Polars sort order), and warns once with their counts. The check stays one collect.
• The kernel reads only the valid rows through a zero-copy slice, with no filter pass.
• A null or NaN in the middle of x, or nulls at the end (for example after sort(nulls_last=True), which keeps a sorted flag), still raises "not sorted".
• The file scan path already dropped these rows. Its output equals the in-memory output.

Integer bounds past the dtype range:

• A bound past the range of an integer column cast to null. On a line, a pan below 0 on a UInt64 x started the window in the null rows, and a pan above 255 on UInt8 emptied the line, also on data without nulls. A zoomed histogram or 2D histogram on such a column, and a range selection on it, returned nothing.
• _typed_range_bounds now clamps a numeric integer bound to the dtype range, and a range wholly outside it is empty. The line windows, the histogram grids and the selection predicates all use it.
• A bound that is not a number (a string or null from a hand-made request) keeps the plain cast it had before.

Measured on a resident 50M-row frame (4 threads, Float64, Float32, Int64 and Datetime x): the data check and the warm requests take the same time as before within noise, the peak memory stays at 0.1 MB (the slice is zero-copy), and a frame with a null prefix and a NaN suffix costs the same as a clean one.

Tests: minmax, lttb and fpcs with a null prefix and a NaN suffix equal the clean frame; scan equals in-memory; a multi-chunk null prefix; one warning per source and column; the sorted-flag cases; line zooms past the dtype range; histogram zooms and range selections past the dtype range equal the Int64 and Float64 results; non-numeric bounds.

Closes #25