support xarray conversion for list-based stress packages

[wpbonelli]

Package.to_xarray() used to raise for list-based stress packages. Support stress period data conversion to an xarray dataset, in one of two forms, dense or sparse.

Dense

The dense form is the default, where each field becomes a (per, ...) variable, laid out along the rows' primary key, which can be:

• cell ID (CHD, WEL, TVK) on the grid dims
• feature number (UZF, SFR ifno) on the referenced table's size dim, e.g. nuzfcells or nreaches. With grid=True it goes on the grid instead, via cellid
• package index (SPC bndno) on a correspondingly named dim

The typical case for this is a variable living on the grid, e.g. CHD

Dimensions:    (per: 3, layer: 1, row: 3, col: 3)
Coordinates:
    boundname  (per, layer, row, col) object 'a' nan nan ... nan
Data variables:
    head       (per, layer, row, col) float64 1.0 2.0 3e+30 ... 3e+30

Details:

• Aux columns: one variable per aux name.
• Array columns: an extra dim named for their size (EVT's pxdp on nseg-1).
• Keystring packages (SFR, LAK, MAW, LKT, ...): one variable per setting. A setting's own index (SFR's idv) is a dim, and named settings (AUXILIARY conc 1.0) are one variable per name.
• Missing positions: filled with FILL_DNODATA, FILL_INT64 or None.
• Periods carry forward as in MF6: a plain period block replaces the previous one, keystring settings each hold until set again, and a period with no block holds the last one.
• Boundnames are a boundname coordinate.
• Time-series values are NaN, with the series name in a <var>_ts coordinate (e.g. head_ts).
• Duplicates: two rows at one position in a period raise. duplicates="sum" | "first" | "last" merges them instead.

Sparse

sparse=True keeps every row, so there are no duplicates to resolve and the round trip is exact. Rows go along an entry dim spanning all periods, with each row's period, key and keystring setting as coordinates, a cellid split by grid dim. The periods attr lists the periods with a block, so an empty block and no block stay distinct. Packages whose rows have no single key (HFB's two cellids, MVR, OC) have only this form. E.g. CHD

Dimensions:       (entry: 3)
Coordinates:
    per           (entry) int64   0 0 2
    cellid_layer  (entry) int64   0 0 0
    cellid_row    (entry) int64   0 0 1
    cellid_col    (entry) int64   0 1 1
    boundname     (entry) object  'a' nan nan
Data variables:
    head          (entry) float64 1.0 2.0 3.0
Attributes:
    periods:  [0 2]

Round trip conversion

Package.from_xarray() inverts either form. From sparse it's exact. From dense, it writes a period block only where a period differs from the previous one, and a keystring row only where a setting changes, so period data become sparse. That gives the same model, not the same rows: row order and settings that repeat a value aren't kept.