Enforce stricter types for formatters in MappedMatrixVis

[axelboc]

Follows #1702

As planned, the problem became much simpler: all the typing complexity related to formatters in MappedMatrixVis can now be moved into the getFormatter utility function, now called getCellFormatter.

I add a type guard called assertNdArrayValue that allows narrowing down an NdArray given an already narrowed-down DType. After this type guard, dataArray.get(row, col) is properly typed and we can pass the value to a strictly typed formatter without any casting.

I'll do something similar in ScalarVis.

---

The broader issue is that every time we narrow down a dataset object (e.g. Dataset<ArrayShape, PrintableType> => Dataset<ArrayShape, EnumType>), we have to narrow down the value right after with assertArrayValue/assertNdArrayValue/assertPrimitiveValue (e.g. ArrayValue<PrintableType> => ArrayValue<EnumType>). This is because we don't store dataset values within dataset objects. Maybe I'll try to tackle this one day.

[axelboc]

Was a bit too quick on this one: it doesn't work at all with printable compound datasets...

Ideally, compound ndarrays would be typed with a tuple (NdArray<[Primitive<Field0Type>, Primitive<Field1Type>, ...]>) so that the return type of compoundNdArray.get(0) would be automatically inferred as Primitive<Field0Type> (which we could pass to a strictly typed formatter: ValueFormatter<Field0Type>).

Unfortunately, this doesn't work: the return type of compoundNdArray.get(0) is always inferred as Primitive<Field0Type> | Primitive<Field1Type> | .... It's a limitation of the ndarray library and I'm not sure it's fixable.

Seems that we're stuck with returning a weakly typed formatter from getFormatter (ValueFormatter<PrintableType>). Anyways, at least the typing weakness no longer leaks into MatrixVis, so it wasn't all worthless.