projection: add ambisonic mapping matrices for 6th to 14th order

[abhayap]

Channel mapping family 3 already accepts up to 227 channels, but the built-in projection matrices stop at 5th order, so anything above that fails in opus_projection_ambisonics_encoder_create(). Add the missing tables for 6th to 14th order in a new src/mapping_matrix_hoa.c.

These are not built the way the 4th- and 5th-order tables are. Those sample the spherical-harmonic matrix at Fliege-Maier cubature nodes and invert it in the least-squares sense, which is fine to 9th order and then degrades badly: cond(A) goes 3.8 at 9th order, 86 at 10th, 723 at 12th, because those node sets are chosen to be good cubature rules rather than good interpolation points, and the 121-point set is the first with a negative weight. Conditioning is what the demixing step costs in SNR -- per-sector coding noise returns ||A||_F.||A^-1||_F/N times louder in the ambisonic domain -- so extending that construction would give up 15 dB at 10th order and 32 dB at 12th.

Instead these sample at maximum-determinant (Fekete) nodes, which come in sets of exactly (N+1)^2 points chosen to maximize |det Y|, and take the polar factor of the result: the nearest matrix with orthonormal columns. That is energy-preserving ambisonic decoding (Zotter, Pomberger & Noisternig, Acta Acustica 98(1), 2012) at the critically sampled limit L = (N+1)^2. At every order it gives cond(A) = 1, needs no gain field, makes the demixing matrix the exact transpose of the mixing matrix, and holds the Q15 round-trip error at the quantization floor instead of letting it grow with order.

Also:

• Replace the per-order if/else chains in opus_projection_encoder.c with mapping_matrix_get_ambisonic(), so adding an order is a table entry rather than four more branches in two places.

• Drop the 65004-octet cap in mapping_matrix_get_size(). That is the most a mapping matrix can be and still fit an Ogg OpusHead, which is a container limit for the muxer to enforce, not an allocator limit; it excludes only 13th and 14th order (78 KB and 103 KB of cells). The 255-channel bound stays.

• Extend tests/test_opus_projection.c: check that each built-in pair inverts itself to within Q15 (46-64 dB, on a diffuse field), and that a synthetic ambisonic signal -- four tonal sources encoded as plane waves in ACN/SN3D -- survives encode and decode at every order, with and without the non-diegetic channels (~27 dB at 64 kbit/s per channel). The codec figure averages two sets of source directions, because how sparsely four sources happen to land on the sectors moves it by 1-2 dB either way, which is more than the matrices differ by. The synthesis is deterministic and self-contained; no test vectors are needed.