This commit is a small followup for PR#426 to handle another case
with the custom unmarshal. This was found in a complex TOML based
image customization "blueprint".

Otherwise round-tripping something like 5e+22 won't work.

lexNumberOrDateStart() is only called from lexValue() on isDigit(r),
so !isDigit(r) in there can never happen.

It has lx.skip(isWhitespace), so the isWhitespace(r) can never be true.

This doesn't really need tests IMO, but easy enough to write. The end
goal is to have 100% test coverage at least for lex.go and parse.go,
which is helpful for finding new test cases for toml-test.

Never referenced; not sure what it was used for in the past, but it's no
longer.

Fixes:

	./encode.go:231:11: non-constant format string in call to (*github.com/BurntSushi/toml.Encoder).wf
	./encode.go:233:11: non-constant format string in call to (*github.com/BurntSushi/toml.Encoder).wf
	./encode.go:282:10: non-constant format string in call to (*github.com/BurntSushi/toml.Encoder).wf
	./encode.go:284:10: non-constant format string in call to (*github.com/BurntSushi/toml.Encoder).wf
	./encode.go:286:10: non-constant format string in call to (*github.com/BurntSushi/toml.Encoder).wf
	./encode.go:300:11: non-constant format string in call to (*github.com/BurntSushi/toml.Encoder).wf
	./encode.go:315:11: non-constant format string in call to (*github.com/BurntSushi/toml.Encoder).wf

Fixes #445

Fixes #454