video_reader-rs_build.yml: add build workflow - #1

…rash diagnostics

All six austinp tests of the functional suite died with SIGSEGV on the runner
while every austin test passed. The core dump, taken with upstream's own
gdb/core-pattern setup, names the line:

    #0  mojo_event_handler__handle_new_frame (frame=0x19fee0) at events.c:83
    #1  event_handler__emit_new_frame (frame=0x19fee0) at src/events.h:135
    #2  _py_thread__unwind_native_frame_stack at py_thread.c:518

events.c:83 is mojo_ref(frame->scope->key), and the scope it reads is
UNKNOWN_SCOPE, the (cached_string_t*)1 sentinel that austin puts in a frame
when libunwind cannot name it. Patch 0006 gives such a frame a real
"<unknown>" cached string instead. The crash-diagnostic steps stay in the
test job: they are what upstream's tests.yml does, and they are what turned
this from an opaque -11 into one line of C.

The inline `sed` PATCH_COMMAND added for deboost.context never ran: with
nle's cmake_minimum_required(3.28) FetchContent populates through a
generated sub-build project, and re-serializing the command into it
expanded ${CMAKE_SYSTEM_PROCESSOR}, stripped the quotes and the `\/`
escapes, so every build failed at configure time with
"sed: -e expression #1, char 19: unknown option to `s'".

Do the edit from a small `cmake -P` script added by the patch instead;
only its path has to survive the sub-build round trip. The script is
idempotent and a no-op if deboost.context gains its own riscv64 branch.

Also set SETUPTOOLS_SCM_PRETEND_VERSION: applying the patch leaves the
tagged checkout dirty, so setuptools_scm would otherwise version the
wheel 1.3.1.dev0+g<sha>.d<date> (the failing runs already reported
"Building nle backend version: 1.3.1").

And fix the test step the build never reached: numpy and psutil only
resolve to riscv64 wheels from pypi.riseproject.dev, and
test_profile.py needs the pytest-benchmark fixture even though it skips
its benchmarks when CI=true (upstream tests against nle[all]).

* Add nle riscv64 wheel build

NetHack Learning Environment: pybind11/CMake extension over the NetHack
game engine (NGPL-licensed), no riscv64 wheel on PyPI (manylinux x86_64/
aarch64 only, cp310-cp313). Mirrors upstream's own cibuildwheel job from
.github/workflows/test_and_deploy.yml, narrowed to riscv64 (dnf instead
of yum for the manylinux2014-vs-Rocky-10 package manager difference).

Patches the FetchContent-fetched septag/deboost.context dependency
(pinned by nle's own CMakeLists.txt) to add the riscv64 branch its own
CMake arch dispatch is missing: it already ships riscv64 fcontext
assembly (asm/*_riscv64_sysv_elf_gas.S) but never selects it, silently
falling through to the x86_64 files instead.

* nle: fix deboost.context patch step, pin wheel version, fix test env

The inline `sed` PATCH_COMMAND added for deboost.context never ran: with
nle's cmake_minimum_required(3.28) FetchContent populates through a
generated sub-build project, and re-serializing the command into it
expanded ${CMAKE_SYSTEM_PROCESSOR}, stripped the quotes and the `\/`
escapes, so every build failed at configure time with
"sed: -e expression #1, char 19: unknown option to `s'".

Do the edit from a small `cmake -P` script added by the patch instead;
only its path has to survive the sub-build round trip. The script is
idempotent and a no-op if deboost.context gains its own riscv64 branch.

Also set SETUPTOOLS_SCM_PRETEND_VERSION: applying the patch leaves the
tagged checkout dirty, so setuptools_scm would otherwise version the
wheel 1.3.1.dev0+g<sha>.d<date> (the failing runs already reported
"Building nle backend version: 1.3.1").

And fix the test step the build never reached: numpy and psutil only
resolve to riscv64 wheels from pypi.riseproject.dev, and
test_profile.py needs the pytest-benchmark fixture even though it skips
its benchmarks when CI=true (upstream tests against nle[all]).