Previously only 64-lead and 40-lead models were identified. This patch
adds 48-lead and 32-lead device identification as well.

Please note that due to missing parameter negotiation functions one
might get the following message with a subsequent crash:

WARNING: from element /GstPipeline:pipeline0/GstV4l2Src:v4l2src0: Could
not get parameters on device '/dev/video0'

This can be avoided by explicitly specifying the format as follows:

gst-launch v4l2src ! deinterlace tff=1 method=4 ! 'video/x-raw-yuv,
width=(int)720, height=(int)576, format=(fourcc)I420' !
nv_omx_videomixer ! nv_gl_eglimagesink

Powering on, improved interlacing.

Use right buffer flag NVMAP_HANDLE_WRITE_COMBINE to allocate buffer,
which can be shared by VI/CSI and CPU. Don't use NVMAP_HEAP_SYSMEM.
It is validated to old T20 silicon and can't support big buffers. By
default, our nvmap_alloc() will use IOVMM to allocate buffers.

nvmap_pin() gives us IOVA for hardware engines like VI/CSI module
with IOMMU enabled in kernel. nvmap_mmap() gives us VA for CPU
read/write operations. So we need to convert VA address to physical
address of the buffer and map that buffer to user space processor's
memory space "page by page".

Change-Id: I543d9d95fc14395200647e09480f25d9bc001e00
Signed-off-by: Bryan Wu <pengw@nvidia.com>

This reverts commit 080c3135b58aa700851991fb672e6c33cf16d9d9.

Fix issue with nvhost_client_device_init being issued more than once.

bug report courtesy of Mikhail Shevchuk