Tests: Drop the GPU kernel-image skip and define one packet test marker

Restore test/gpu_common.h: a binary without a kernel image for the device
fails again instead of exiting 77, which CI counts as a pass. Packet test
kernels and operations share EIGEN_PACKET_TEST_FUNC from
packetmath_test_shared.h.

Correct the comments on the rsqrt double budget under MSVC, the uint8_t
operand range and the alignment the aligned device loads rely on, and drop
an unused template parameter.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
4 files changed