This is a really cool approach! But if it so obvious, why doesn't every hashmap use it? It seems like there are some trade-offs here that I must be missing.
If you do any kind of probing other than linear probing (e.g. quadratic probing) this approach doesn't work anymore, because your collisions are no longer densely grouped together.
The main tradeoff is concurrency: it's difficult to safely read a hash table while its entries are being concurrently relocated. Another tradeoff is (probably) higher average latency (but worst-case latency is much better since there's no global rehashing required).