wrt optimize:  how much time is spent generating the noise vs drawing it?  not much you can do to optimize 256K * 4 calls to noise, but other stuff to try:
  use P2D instead of JAVA2D for faster pixels[]
  use System.arraycopy() instead of your manual pixel-copy loops
  allocate the tile array ONCE and reuse
  bitwise assemble the rgb into tile[] rather than calling color()

wrt seamless:
  any continuous fn that wraps on uv [0..1] will work.  think about building a table of trig values for a full cycle, then indexing into it by integer multiples... for instance.

hth