Improve number handling by returning NaN for non-numbers

[DocLambda]

When processing unknown documents, fields that should contain numbers might, for whatever reason, actually contain e.g. a string. In this case the current implementation is also a bit inconsistent and will return a float64(0) when performing a number(//some_string_field) query and will panic if executed on a string. The latter might happen if the query is constructed prgramatically.

The present PR changes the behavior such that the above query will return math.NaN() and leaves the value handling to the caller. If required, the previous behavior can be reconstructed in ceil, floor etc which will now also return NaNs.