Use demand in vector.read

[jamii]

Hooking up demand to vector.Read.

• Use demand in vector.Read to avoid reading unneeded vectors. Also changes the type of the output value, removing any unused fields so that we can avoid having to materialize a bunch of nulls.
• Add vngio.ReaderOpts, where opts.Demand defaults to demand.All{}.
• Split the demand representation into a separate package to avoid import loops.
• Pull the fuzzer code from Fuzz vng roundtrip #4833 into a fuzz package. (I've manually copied across the requested changes from that pr).
• Add a fuzzer that runs random queries through both zng and vng paths and compares the results. (Found value.Under loops forever if called with null union #4854). This uses a gross hack to get around the fact that readers are created before queries are analyzed.
• Change the constants vector to hold []byte instead of zed.Value, since all we ever do is copy the bytes out.

This doesn't yet actually do anything for zq or zed query. I can work on the latter, and @nwt is thinking about how to refactor to make zq possible.

[jamii]

@brimdata/zed This is ready for review, but before merging I'll need to rebase it on #4833 to make the vng tests use the new fuzz package.

[jamii]

@nwt I finished rebasing this on the other two PRs.

[nwt]

I'd like to have the NewReader(zctx *zed.Context, r io.Reader, d demand.Demand) change but then .

[nwt]

Nits:

1. Could use io.Writer here to make this a little more flexible.
2. We usually put the destination parameter first (as with zio.Copy).
3. We usually use val as the name for zed.Value slice.

Suggested change

	func WriteZNG(t *testing.T, valuesIn []zed.Value, buf *bytes.Buffer) {
	func WriteZNG(t *testing.T, w io.Writer, vals []zed.Value) {