SPIR-V is on a completely different level. The Vulcan API is called by the binary running on the CPU. The binary generated from SPIR-V is executed on the compute device, which in 99% or all cases that Vulkan is concerned with will be the GPU.
In the same way you don't use GLSL to program OpenGL, you don't use SPIR-V to program Vulkan. GLSL/SPIR are to OpenGL/Vulkan what browser-side JavaScript is to a webserver.
Sure, but like PTX/CUDA it will mean you can write shaders in other languages with a better hardware mapping, instead of compiling them to GLSL/OpenCL C source, which is currently one advantage of CUDA.
Oh yes, I've been waiting to do that with OpenGL for a looong time. As a matter of fact I preferred working with the ARB_vertex_program and ARB_fragment_program extensions; I even designed (but never fully implemented) a custom, Lisp inspired language that compiles into the ARB_???_program assembly.
In the same way you don't use GLSL to program OpenGL, you don't use SPIR-V to program Vulkan. GLSL/SPIR are to OpenGL/Vulkan what browser-side JavaScript is to a webserver.