**BREAKING** pygmt.grdcut: Refactor to store output in virtualfiles for grids

[seisman]

This PR applies virtualfiles in pygmt.grdcut, so that this function doesn't have to write to temporary files (#2730).

grdcut is on of the complicated module that it works for grid, image and cube. We have to know the kind of the input before calling the Session.virtualfile_out method.

It turns out it's not trivial to determine the data kind of an input file. We have some options:

1. Based on file extensions (e.g., .nc is a grid, .tif is a image)
2. Based on magic bytes (xref: https://en.wikipedia.org/wiki/List_of_file_signatures)
3. Try reading as a grid/image/cube. If fail, then try the 2nd one.
4. Add a new kind parameter (default to "grid") and users should specify the kind parameter.

Option 1-2 may not always work; option 3 also dosn't always work (see #3115 (comment)).

This PR currently adopts option 4. Need to note option 4 meaning BREAKING:

• previously, grdcut("@earth_day_01d") works; now must use grdcut("@earth_day_01d", kind="image")
• previously, grdcut works 1/3/4-band images; now it only supports 3-band images. [We can support 1/4-band images later]

[seisman]

I think it's better to not introduce last-minute breaking changes, so bump this PR to v0.15.0.

[weiji14]

Still think using the private gmt_raster_type func as mentioned at #3115 (comment) would be the best solution to auto-detect the raster kind, but failing that, the logic here to explicitly have the user pass in the type works too.

Just one typo on the match-case statement, and should be ok to merge once all tests pass without coverage misses.