msm: bam_dmux: modify DL low memory algorithm

Bam_dmux maintains a pool of dynamically allocated buffers for the BAM DMA
engine to use as destination memory for Downlink (DL) data.  The allocation
requests for the buffers are done using flags to indicate delays are
tolerable and either the memory or an error should be returned immediately
to bam_dmux as the buffers are normally allocated within various hot paths
inside of the driver.  If a low memory allocation failure is encountered
when attempting to replenish the pool of buffers, bam_dmux will either
attempt the memory allocations later when the next scheduled pool refilling
occurs if there is at least one buffer remaining in the pool, or will
schedule a delayed task to retry the buffer allocation until atleast one
buffer can be put into the pool.

This mechanism of handling low memory situations has two flaws.  First, the
allocation type used by bam_dmux does not trigger the standard low memory
recovery mechanism within the memory management subsystem.  Second, the
peripeheral on the otherside do the DMA engine has a limited hardware based
flow control mechanism that will not release flow control with only one
buffer queued to the DMA engine.

To assist in system recovery, and avoid a data stall, create a cold path
that is tolerant of memory allocation delays and is triggered when the hot
path encounters an allocation failure.

CRs-Fixed: 589665
Change-Id: If72b0eaef285ac6dfd99e97bef1cd7df154e3a48
Signed-off-by: Jeffrey Hugo <jhugo@codeaurora.org>