Do not reserve memory for the bucket count read from a time series state

The number of buckets is read from the serialized state and was only checked against
the bucket count derived from the function parameters. A huge window makes that count
enormous, so a tiny state could ask for an arbitrarily large allocation: an 18-byte
state of `timeSeriesChangesToGrid` allocated 1.25 GiB. Only a bounded number of buckets
is reserved now and the map grows while the buckets are read.