Fill on init
Zeros are a special case of a number. initCpu grows an optional fill: ?T. No initial data, you memset the buffer to the fill, or to 0. A test asks for a {2,3} of twos and checks the slice. PR #10 — “1 create tensorfill to fill with value” — merges it.
This is still CPU. initCpu still takes a device_id you could set to GPU and then fail later in slice. Fill is part of construction, not a method you call on a tensor that already exists. That method is a year away.