Add reinterpret as interface to bitcast, pack, and unpack

[AntonOresten]

Adds Base.reinterpret(::Type{T}, ::Tile) and Base.reinterpret(reshape, ::Type{T}, ::Tile), giving tiles the same whole-array reinterpret semantics as AbstractArray: the bits are viewed as one contiguous column-major block and the leading dimension is rescaled by the element-width ratio:

julia> arr = rand(UInt8, 2, 2)
2×2 Matrix{UInt8}:
 0x11  0x13
 0xfd  0x20

julia> reinterpret(UInt16, arr)
1×2 reinterpret(UInt16, ::Matrix{UInt8}):
 0xfd11  0x2013

julia> reinterpret(reshape, UInt16, arr)
2-element reinterpret(reshape, UInt16, ::Matrix{UInt8}) with eltype UInt16:
 0xfd11
 0x2013

This is distinct from the element-wise reinterpret.(T, x) broadcast, which preserves shape and already lowered to bitcast.

Here's FP4, loaded two-per-byte:

bytes = ct.load(a, pid, (8,))             # Tile{UInt8,(8,)}
fp4   = reinterpret(Float4_E2M1FN, bytes) # Tile{Float4_E2M1FN,(16,)}
vals  = convert(ct.Tile{Float32}, fp4)

I think this is justified since Tile lacks a byte-level memory representation, and Tile IR treats sub-byte types like Float4_E2M1FN as first-class tile elements. Their sub-byte-ness only surfaces at the memory boundary, which pack/unpack bridge.

Lowering

With q = bitwidth(eltype(x)) and p = bitwidth(T), the leading dimension always scales by q // p; the op is chosen by which side is the byte type (i8):

condition	op
p == q	bitcast (shape preserved)
target is a byte (p == 8)	pack — to bytes
source is a byte (q == 8)	unpack — from bytes
neither	pack then unpack (pivot through i8)

pack/unpack (new intrinsics over cuda_tile.pack/unpack, v13.3+) are rank-1 ops between a numeric tile and a byte tile, named for their direction relative to bytes — not for element-count growth, which the q ÷ p scale carries (a sub-byte unpack expands the dimension; a super-byte pack like UInt16 → UInt8 also expands it). reinterpret flattens to rank-1 via reshape, converts, and reshapes back, so it works at any rank; identity reshapes fold away, leaving a single pack/unpack in the common case. Mirrors cutile-python's pack_to_bytes/unpack_from_bytes.

bitwidth trait

reinterpret on sub-byte types needs a true element width, since sizeof rounds up to whole bytes (Float4_E2M1FN is 4 bits, sizeof == 1). Adds cuTile.bitwidth(::Type), defaulting to 8 * sizeof(T); the Microfloats extension forwards to Microfloats.bitwidth.