Refactor gcsfs to list file contents via prefixed bucket listing,
rather than cached exhaustive bucket listing. In progress, but provides
basic interface compatibility for walk, glob, ls, info. Intended to
support re-addition of metadata caching via the _list_objects interface
to provide prefix-specific listing caches.

Update `info` to retrieve object info via object get.

Add per-directory listing cache to GCSFS caching object metadata under
the given directory. Resolves listing requests via cache, supporting
walk/ls/glob/etc. Resolve `info` requests via cache if the parent
directory has been listed, otherwise directly request object data.
Updates cache invalidation logic to function on path prefixes, allowing
object writes to invalidate their parent/sibling caches, rather than
entire listing cache.

Add decorator-based method tracing to `gcsfuse.GCSFS` and
`core.GCSFileSystem` interface methods. Add `--verbose` command-line
option to `gcsfuse` to support debug logging control.

Prototype `per_dir_cache` integration for gcsfuse. Minimal fixup to
gcsfuse to support directory listing.

Fix error in GCSFS::read() cache key resolution.

Resolve error when writing small partitions via dask.bag.to_textfiles.
Error occurs when partition size is below minimum GCS multipart upload
size.

Close logic in dask.bytes.core calls flush(force=False), followed by
flush(force=True) on GCSFile. Current logic initializes multipart upload
on non-force flush and attempts to write a non-final block below the
minimum GCS upload block size.

Fixup logic to skip flush if buffer size is below minimum upload size on
non-forced flush. This, incidentally, avoids initialization of multipart
upload in cases where final file size will be below the minimum block
size, which was resulting in duplicate uploads for small output
partitions.

Add tracing logic to GCSFile file operations for debugging. Update
`_tracemethod` to perform, optional, traceback logging at `DEBUG-1` log
level.

Updates `ls` to return non-prefix separated prefix search, needs to be
verified? Should this be glob-like? Fix error from dask.bytes when
read-only file is flushed. Fixup returning listing with "path"
attribute.

Retry on requests failing due to `google.auth.exceptions.RefreshError`,
partial resolution of fsspec#71.

Resolve error when writing small partitions via dask.bag.to_textfiles
when partition size is below minimum GCS multipart upload
size.

Close logic in dask.bytes.core calls flush(force=False), followed by
flush(force=True) on GCSFile. Current logic initializes multipart upload
on non-force flush and then attempts to write a non-final block
below the minimum GCS upload block size.

Fixup logic to skip flush if buffer size is below minimum upload size on
non-forced flush and instead issue a warning. This, incidentally, avoids
initialization of multipart upload in cases where final file size will
be below the minimum block size, which was resulting in duplicate
uploads for small output partitions.

Update core.py to lift GCS block size limits into module level
constants. Replace use of constants in core.py with symbolic names.

From fsspec#73 review. Defer multipart upload if a simple upload may be at the
specified block size on non-forced flush. Minor reorganization of
`flush` logic to group error handling vs deferral.

Relax block size restrictions on fetch, no longer aligning `range`-ed
fetch requests to block boundaries.

Fix minor logging error in `_fetch`.

Updates GCSFileSystem cache configuration. Set cache as non-expiring in
default configuration, but continue to allow configurable cache timeout.
Do not bypass cache on `_get_object` calls if object is not present in
cache listing.

Updates GCSFileSystem docstring to include description of caching
semantics.

Adds explict flag to control stacktrace debugging for traced methods.
Reduces log size on test failures.

From review, cleanup `walk` implementation. Fix pseudodir creation on
bucket-level `info` call. Remove `norm_path` todo.

`flush` on an open, but read-only, file should be a no-op, not raise
a ValueError, compare to  builtin `open("read_only", "r").flush()`.
Updates `flush` logic and adds test case covering flush behavior.

…ache