[docker] bump ROCm

[lauri9]

Bump ROCm to version 10.1.0 (nightly).

Closes #954

Summary of changes

Bump ROCm stack.

Two workarounds are needed:

• libsqlite3-dev is installed to workaround rocprofiler-sdk SQLite header gap - [Issue]: rocprofiler-sdk SQLite include path relies on incidental manylinux header ROCm/TheRock#5868
• Install tcl to work around profiler-hub SQLite amalgamation build failure - [Issue] rocprofiler-systems configure fails on build hosts without Tcl: profiler-hub's bundled SQLite3 amalgamation requires tclsh ROCm/rocm-systems#10059

Furthermore, the previous installation method for amdsmi Python package no longer works and is updated.

Tasks

• Bump ROCm, verify build works
• Update MIOpen tuning databases
• Verify performance and visual quality on gfx950
• Verify performance and visual quality on gfx942
• Verify profiling works

Performance

CI run MI350X is OK, checked a couple of configs manually. No regressions, visual quality good.

model	ROCm bump	baseline main abeb76a	Ratio
flux.usp	0.76	0.78	102.5%
flux.single_gpu	2.27	2.33	102.7%
flux.usp_2k	2.13	2.14	100.4%
flux.single_gpu_2k	11.68	11.81	101.1%
flux2.default	4.24	4.23	99.7%
flux2.quantgemm.gfx950	2.33	2.35	101.0%
flux2.quantgemm_quantattn.gfx950	2.15	2.20	102.0%
flux2.ti2i_2k	12.35	12.45	100.8%
flux2.t2i_1k	2.75	2.81	102.0%
flux2.t2i_2k	10.43	10.59	101.6%
flux2.t-multi-i2i_1k	5.64	5.62	99.5%
flux2.t-multi-i2i_2k	14.93	14.99	100.4%
flux2_klein.default	0.47	0.47	100.9%
flux_kontext.default	1.44	1.45	100.7%
hunyuanvideo.default	102.41	104.26	101.8%
hunyuanvideo_1_5.default	115.83	115.46	99.7%
ltx2_5.default	7.37	7.38	100.1%
minimax_h3.default	32.80	32.99	100.6%
qwen_image.default	4.74	4.73	99.6%
qwen_image_edit.default	14.54	14.75	101.4%
stablediffusion_3_5.default	0.82	0.82	100.9%
wan2_1.default	73.00	73.10	100.1%
wan2_2.default	74.62	74.20	99.4%
wan2_2.quantgemm.gfx950	66.28	66.45	100.3%
wan2_2.quantgemm_quantattn.gfx950	38.92	38.72	99.5%
wan2_2.quantgemm_sparseattn.gfx950	29.58	29.51	99.7%
z_image.default	2.52	2.55	101.4%

MI300X testing was done on different node than usual, so results are not directly comparable. However, tested on that same node against baseline main commit abeb76a and there were no regressions.

Other

THEROCK_COMMIT is the last commit before the compiler update that triggered ROCm/llvm-project#4675