This should fix some arch-specific problems with the "generic DMA"
API implementations of dma_supported() and dma_set_mask():

  - dma_set_mask():
       * Removes pointless clones that were used on arm, i386,
         ia64, mips, and parisc
       * Also removed PCI-specific "asm-generic" version
       * Replaces them with a truly generic implementation that
         doesn't needlessly BUG() with non-PCI devices
       * Adds an ARCH_HAS_* hook to cope with sparc64-only
         restrictions (for ali5451 pci audio hardware);
         sparc64 retains a BUG() for non-pci hardware.
  - dma_supported():
       * Removes broken versions from arm, i386, mips; the
         broken versions never actually compared the two
         masks they were specified to compare:
            + arm only verified there was a device mask
            + i386/mips didn't even do that much
       * Also removed PCI-specific "asm-generic" version
       * Replaces them with a generic implementation that
            + doesn't needlessly BUG() with non-PCI devices
            + actually compares the two masks (bit-wise)
       * Adds an ARCH_HAS_* hook to handle ia64 and parisc,
         which already have their own fancy implementations

Sanity tested on x86/pc and ARM/PXA.

It's unfortunate that the arch-specific BUG() behavior only
started getting reported recently (test7?); and particularly
that the previous dma_supported() was so broken even on i386.
In consequence, it's not quite clear what bugs are being
hidden by the bogus masking.

- Dave