Update dlpack to v1.0rc (#1667)

* Implements DLPack changes necessitated by memory object work

Use `ary_base` rather than `usm_ary` as manager_ctx in
     DLManagedTensor struct.

Tests for DLPack use fixture to provide all devices to take advantage
of caching.

* `usm_ndarray` constructor now honors the writable flag when passed a read-only array as `buffer` argument

* Adds support for DLPack v1.0

* Uses `dlpack.h` as per DLPack v1.0 rc
* Implements `device` and `copy` keywords into `__dlpack__` and `from_dlpack`
* Implements `max_version` argument into `__dlpack__`
* Adds `DLDeviceType` enumerator to `_usmarray.pyx` for representing device types as enumerators
* Reimplements `dpt._dlpack.get_build_dlpack_version` to return a tuple of the major version and the minor version

* Adds check for major version of versioned dlpack tensor to `from_dlpack_versioned_capsule`

Also adjusts some error messages in _dlpack.pyx

* Add missing device_type enumerators from `dlpack.h`

* usm_ary -> ary_base in `to_dlpack_versioned_capsule`

* Changes to exception messages throughout _dlpack.pyx

* Versioned DLpack tensor deleter and owner tweaked

These tweaks are based on similar changes made for the managed tensor during memory work

* Updates `dpctl/tensor/include/dlpack/README.md`

DLPack version updated and new SHA for `dlpack.h` added

* Adds check that `max_version` is a 2-tuple to `__dlpack__`

* Add more dlpack tests (#1670)

* Fixture all_root_devices is to keep only 2 devices from each platform

This is done to speed-up test suite execution on multi-GPU system,
like Aurora.

* Add tests for legacy/versions capsule import-export

* Expand tests more to improve coverage

Use F-contiguous array to both legacy and versioned capsule.
Add tests for read-only array and for use of copy=True keyword
for __dlpack__ call

* Test the writable flag of a usm_ndarray from read-only DLPack capsule

Also adds a test checking for the writable flag of a usm_ndarray constructed with a read-only usm_ndarray as the buffer

Removes some commented out `flags` checks from `test_meshgrid2`. These checks were malformed, as `dpt.meshgrid` returns a non-contiguous view in the test.

* Factor out device-id computation

Reuse that in both capsule producers to avoid code duplication

* Reuse get parent device ordinal id routine (#1672)

* Reused get_parent_device_ordinal_id routine

* test_legacy_dlpack_capsule uses 4 kinds of dtype

Added test to use non-default copy keyword, and non-default device keyword
argument.

* Adds more DLPack tests to improve coverage

* Adds a test for boolean arrays in test_legacy_dlpack_capsule

* DLPack tests for `copy` and `stream` kwargs in `__dlpack__`

---------

Co-authored-by: Oleksandr Pavlyk <[email protected]>

Author: ndgrigorian

dpctl/tensor/_dlpack.pxd

@@ -24,15 +24,24 @@ from ._usmarray cimport usm_ndarray
 
 cdef extern from 'dlpack/dlpack.h' nogil:
     int device_CPU 'kDLCPU'
-    int device_oneAPI 'kDLOneAPI'
+    int device_CUDA 'kDLCUDA'
+    int device_CUDAHost 'kDLCUDAHost'
+    int device_CUDAManaged 'kDLCUDAManaged'
+    int device_DLROCM 'kDLROCM'
+    int device_ROCMHost 'kDLROCMHost'
     int device_OpenCL 'kDLOpenCL'
-
+    int device_Vulkan 'kDLVulkan'
+    int device_Metal 'kDLMetal'
+    int device_VPI 'kDLVPI'
+    int device_OneAPI 'kDLOneAPI'
+    int device_WebGPU 'kDLWebGPU'
+    int device_Hexagon 'kDLHexagon'
+    int device_MAIA 'kDLMAIA'
 
 cpdef object to_dlpack_capsule(usm_ndarray array) except +
+cpdef object to_dlpack_versioned_capsule(usm_ndarray array, bint copied) except +
 cpdef usm_ndarray from_dlpack_capsule(object dltensor) except +
 
-cpdef from_dlpack(array)
-
 cdef int get_parent_device_ordinal_id(SyclDevice dev) except *
 
 cdef class DLPackCreationError(Exception):