add BigInt support to the normalize transform

[Fil]

Note: it's a bit of a bummer in that we're adding overhead for all Arrays (by making a copy of the channel to a Float64Array). Maybe we could just do it if the first non-nullish value is BigInt?

cc: @juba

closes #2154

[Fil]

It should definitely be earlier — in my code we coerce the whole channel as many times as there are series. But where? Maybe as high as in the map transform, where we do https://github.com/observablehq/plot/blob/0032c11160e6340ce44307675d93bed382d10446/src/transforms/map.js#L36

In that case the mapper should possibly have a type key, like so:

  {
    mapIndex(I, S, T) {
      const b = basis(I, S);
      for (const i of I) {
        T[i] = S[i] === null ? NaN : S[i] / b;
      }
    },
    type: Float64Array
  }

that would be passed as the third argument to valueof?

The alternative is (what you say) to preemptively convert all BigInt independently of where they're used.

[Fil]

This is very close to https://github.com/observablehq/plot/issues/473